ICEIS 2026 Abstracts

Area 1 - Databases and Information Systems Integration

Full Papers

Paper Nr:	18
Title:	Markov Chain and Machine Learning Based Approach for Identifying Abnormal Behaviors of Subscribers in Mobile Networks
Authors:	Quang-Vinh Tran, Duc-Thinh Vu, Ngoc-Toan Dinh, Tuan-Bach Phan and Quang-Diep Pham
Abstract:	Detecting individual behavioral anomalies is a critical task to ensure both system security and overall quality of service. In mobile communication networks, the behavior of each user equipment (UE) carries valuable information that can reveal abnormal patterns. Such anomalies may result from device malfunctions that disrupt service for individual subscribers, active attacks initiated by subscribers that compromise operational security, or inefficiencies in current network planning and management. In this work, we focus on rich event log data collected from the live mobile network of Viettel Group, which provides a comprehensive view of subscriber activities across the 5G network system. We transform subscriber behaviors into Markov sequences and apply a Variational Autoencoder (VAE) to detect anomalous cycles that deviate from normal behavioral patterns. To further interpret the detected anomalies, we perform clustering on the anomalous samples to group similar fault behaviors. This two-stage approach-detecting anomalies first and clustering afterward-effectively separates different fault types and mitigates the noise from normal behaviors. By doing so, our approach uncovers hidden issues in real time without the need for manual reporting or intervention from traditional monitoring systems. The findings demonstrate that combining Markov modeling with machine learning enables fast, proactive, and scalable anomaly detection. Ultimately, this work contributes to enhancing the automation capabilities of Viettel’s network operations and improving the quality of service delivered to its 5G subscribers.
Download

Paper Nr:	74
Title:	A Metric-Based Framework for Assessing Semantic Interoperability, Scalability, and Temporal Quality in Energy Digital Twins
Authors:	Amina Sghiri, Marie Gallab, Safae Merzouk and Saliha Assoul
Abstract:	Energy Digital Twins rely on the integration of heterogeneous building data such as Building Information Modeling (BIM) models, Internet of Things (IoT) sensors, Building Management Systems (BMS), and weather information. However, the quality of these integrated systems is rarely evaluated using unified and measurable criteria. This paper proposes a structured evaluation framework for Energy Digital Twins based on three complementary dimensions: semantic interoperability, scalability, and temporal quality. The framework relies on the semantic integration of building information using Industry Foundation Classes (IFC) models, the Brick ontology, and Resource Description Framework (RDF) knowledge graphs, enabling the definition of quantitative indicators for each dimension. The approach is illustrated through a case study of a training room digital twin integrating simulated sensor data and external weather information. The resulting RDF graph contains 612 structural triples and more than 69,000 triples after the integration of time-series observations. The evaluation results show a semantic interoperability score of 0.79, a scalability indicator of 4.23 × 10⁻⁵ seconds per triple, and a temporal quality value close to 1.0 under the simulated conditions. These results illustrate the applicability of the proposed metrics for assessing the quality of Energy Digital Twins and provide a methodological basis for objectively comparing implementations and supporting the design of more reliable and scalable systems.
Download

Paper Nr:	136
Title:	Towards Event Data Analysis: An End-to-End Architecture with a Declarative Query System
Authors:	Felipe F. Vasconcelos, Guilherme M. Rocha, Ana Paula Sodré, Mirian Halfeld-Ferrari, Carmem S. Hara and Cristina D. Aguiar
Abstract:	This article focuses on analyzing event data, which is prevalent in fields as diverse as public health or social media. By extracting and structuring event data from raw sources, insights can be gained through analysis. The paper presents an end-to-end architecture, detailing the process from data extraction to query submission and results visualization. The aim is to show how to build an intuitive, turnkey solution for professionals in different domains, enabling them to derive insights without being burdened by complexity. We highlight the development of interfaces for querying data and visualizing results, together with reliable background tools to ensure accurate and understandable results. To achieve this, the paper introduces a SQL-inspired declarative query system and demonstrates its applicability through a case study.
Download

Paper Nr:	156
Title:	An Integrated Data-Driven Approach to Understanding Unforeseen Traffic Events
Authors:	Michelle Koch, Jhonny Pincay, Damian Nomura and Luis Terán
Abstract:	Unpredictable traffic congestion presents significant challenges to both mobility and logistics in Switzerland. Despite its substantial impact on commuters and supply chains, the factors that influence the duration of non-recurrent congestion-particularly those caused by accidents-remain to be insufficiently understood. This study offers a data-driven investigation of accident-induced traffic delays, focusing on highways in Eastern Switzerland and examining the influence of key variables, including accident severity, weather conditions, proximity to construction zones, and prevailing traffic conditions. To account for the heterogeneity and non-normal distribution of traffic data, robust statistical techniques, including non-parametric methods, are employed to analyze a diverse set of traffic event records. The results evidence that accident severity is the most significant predictor of congestion duration, with severe incidents leading to a 120% increase in median delay. Contrary to prior assumptions, weather conditions did not exhibit a statistically significant influence on delay duration. Construction zones were associated with shorter delays, and 81% of recorded accidents occurred under normal, uncongested traffic conditions, suggesting limited feedback between prior congestion and incident occurrence. These findings refine the understanding of non-recurrent congestion dynamics and underscore the central role of severity-aware modeling in traffic delay prediction. By combining a scalable data infrastructure with rigorous statistical analysis, this work provides an empirical foundation for developing multi-factor congestion prediction models to support more effective traffic management strategies.
Download

Paper Nr:	180
Title:	Auruns: Agent for Unified Resource Utilization in Unified Namespace System
Authors:	Victor Moisés Silveira Santos, Gabriel Mac'Hamilton Renaux Alves and Alexandre Magno Andrade Maciel
Abstract:	Industry 4.0 requires accessible equipment data management across heterogeneous industrial systems. Manufacturing environments maintain equipment information in spreadsheets and databases organized according to Unified Namespace (UNS) hierarchies, but querying this data requires technical expertise or navigation through complex schemas. Existing solutions rely on SQL queries or generic search functions that do not leverage industrial terminology and hierarchical relationships. This paper presents a conversational agent enabling natural language queries over industrial equipment organized according to UNS principles, combining Retrieval Augmented Generation (RAG) with automated hierarchy generation and user interaction analytics. The open-source technology stack (LangChain, Chroma, FastAPI) enables practitioners to implement similar systems for improving maintenance efficiency and production uptime in Industry 4.0 environments. Validated using 1,418 equipment records from an automotive manufacturing plant in Brazil, the system achieved 100% retrieval success rate and zero hallucination (faithfulness: 5.0/5.0) using LLM-as-a-Judge evaluation. The RAG architecture handles technical terminology and multi-level hierarchies with average retrieval latency of 129.73 ms. Evaluation revealed overall quality of 4.87/5.0 across relevance, technical accuracy, grounding, clarity, and faithfulness criteria, with performance consistency across production areas (4.82-4.93) demonstrating robust generalization.
Download

Paper Nr:	193
Title:	Process Framework and Big Data Platform for Evidence-Based Public Management in Health
Authors:	Davyson S. Ribeiro, Victoria T. Oliveira, Rossana M. C. Andrade, Pedro A. M. Oliveira, Ícaro S. de Oliveira, Wilson Castro, Paulo Fabricio, Leonan Carneiro, Tales P. Nogueira and Ismayle S. Santos
Abstract:	A Smart City involves a wide range of participants, from public managers to citizens, and its effectiveness depends on the integration of different sectors and the sharing of information supported by Information and Communication Technologies (ICT). The widespread use of these technologies generates vast amounts of digital data - Big Data - which, when properly managed, can guide more effective public policies. In this dynamic context, it becomes essential for public managers to understand the impacts of their decisions through systematic data analysis. This article presents the “BigData Fortaleza” platform and its process framework, developed to support evidence-based public management in health and urban planning. The framework establishes a structured workflow that connects public managers and data scientists through well-defined stages including problem definition, data acquisition, pre-processing, business logic implementation, validation, and dissemination of analyses. This process ensures data quality, analytical transparency, and compliance with privacy and ethical standards such as the Brazilian General Data Protection Law (LGPD). By integrating Big Data and Artificial Intelligence techniques, the platform enables exploratory, diagnostic, and predictive analyses to address key public challenges identified in collaboration with managers from the Fortaleza City Hall. These analyses have provided actionable insights for developing more effective and sustainable public policies. The proposed framework demonstrates how combining a robust analytical process with advanced technological infrastructure can transform urban governance, offering continuous, data-driven support for strategic decision-making and contributing to the efficiency and transparency of public management.
Download

Paper Nr:	211
Title:	Ontology-Driven Meta-Learning for Adaptive Knowledge Reinjection in CWM-Based Complex Data Warehouses
Authors:	Fabrice Razafindraibe, Jean Christian Ralaivao, Angelo Raherinirina and Hasina Rakotonirainy
Abstract:	This paper proposes an ontology-driven meta-learning framework for adaptive knowledge reinjection in CWM-based complex data warehouses (CDW). Although previous extensions of CWM integrated knowledge, metadata, and learning with reinjection policies, most existing approaches remain strategy-fixed and fail to adapt to evolving domains. To overcome this limitation, the paper introduces a reflexive layer that selects and adjusts learning strategies (inductive, deductive, hybrid) and reinjection policies based on ontological evaluation criteria. The proposed framework aims to improve semantic consistency, robustness to domain shifts, and the operational usefulness of reinjected knowledge. A conceptual use case is used to illustrate the main stages of the proposed lifecycle and its expected benefits. The framework provides a foundation for adaptive and explainable governance of learning in CDW.

Paper Nr:	239
Title:	Causal Discovery and Community Analysis for Detecting Product Interactions in Retail Pricing
Authors:	Jean-Christophe Ricklin, Ines Ben Amor and Raid Mansi
Abstract:	Understanding how price reductions on one product affect the sales of others is essential for retail assortment management: without capturing these inter-product spillovers, the true impact of a price change cannot be assessed. The dominant approach relies on transactional basket data to identify substitutes and complements through co-purchase patterns, but such data can be unavailable due to privacy regulations and aggregated reporting. Alternative approaches based on statistical associations between sales time series can capture co-movements without basket data, but they cannot distinguish genuine causal effects from correlated demand fluctuations, leading to overestimated cannibalization and unreliable control groups. To address these limitations, we propose a graph-based methodology comprising four stages: (1) causal discovery using Markov boundary estimation to construct directed product interaction graphs and equivalence classes from daily sales time series; (2) stabilized community detection via MC-Louvain, a consensus clustering framework that produces product partitions; (3) community-based control group construction, where extra-community products serve as approximately conditionally independent counterfactuals for synthetic control estimation; and (4) comparative evaluation against a statistical promotion effectiveness framework combining association metrics with CausalImpact. The key insight is that community structure serves a dual purpose: extra-community products provide approximately independent controls for estimating direct promotional effects, while intra-community causal links enable the decomposition of spillovers into cannibalization and halo components. Experimental evaluation on the M5 competition dataset shows that leveraging causal community structure enables a substantial improvement in promotion-level ROI estimation compared to association-based statistical approaches.
Download

Paper Nr:	250
Title:	Building Blocks for an RDBMS-only Architecture: A Strategy Pattern-Based Approach for Dynamic Plugin Generation in Data Logic Code
Authors:	Alfonso Vicente and Ariel Sabiguero
Abstract:	Database-centric architectures implement business operations directly within the relational database management system through typed database routines. While this improves encapsulation, invariant enforcement, and architectural cohesion, it also tends to produce repetitive implementations of structurally similar Data Logic (DL) operations across multiple relational contexts. This paper proposes a plugin-based extensibility mechanism for the DL Code layer based on the Strategy design pattern. Generator functions are registered as plugins, declarative bindings associate them with relational contexts, and a regeneration mechanism materializes typed DL routines inside the database. A PostgreSQL-based prototype with multiple plugins shows reduced handwritten DL code, deterministic regeneration under schema evolution, controlled routine exposure for governance, and no observable runtime overhead attributable to an additional execution layer. The proposal is intended for recurrent families of operations, such as range filters, pattern-based searches, and aggregations, rather than for universal synthesis of all possible DL routines. These results indicate that Strategy-based plugin generation is a viable architectural building block for database-centric enterprise systems.
Download

Paper Nr:	270
Title:	Empowering Large-Scale Spatial Analytics in Smart Cities with a Fog-Cloud Lakehouse Architecture
Authors:	João Paulo Clarindo, Rafael Luciano L. Silva, José Matheus S. Alves, João Pedro C. Castro, Fábio J. Coutinho and Cristina D. Aguiar
Abstract:	The exponential growth of IoT-generated spatial data in smart cities challenges existing data management architectures. Traditional data warehouses and data lakes either lack flexibility or fail to provide the performance required for complex spatial analytics. In this paper, we present FlowCaLiSe, a fog-cloud spatial lakehouse architecture designed to support scalable spatial analytics in smart cities. The architecture enables the efficient handling of large-scale spatial datasets, providing scalability, low-latency responses, and advanced analytical capabilities. We also define a structured deployment process to guide the architecture’s implementation. FlowCaLiSe was evaluated through a case study using real-world smart city data and evaluated in a fog–cloud simulation environment, where we analyzed architectural trade-offs in terms of energy consumption and network usage. The results demonstrated the feasibility of FlowCaLiSe for large-scale urban spatial analytics.
Download

Paper Nr:	290
Title:	User Requirements and Affective Drivers of Satisfaction in Women’s Health Apps
Authors:	Azadeh Esfandyari
Abstract:	Women’s health applications represent a rapidly growing segment of the mobile health (mHealth) ecosystem, yet large-scale empirical evidence on user requirements and satisfaction determinants remains limited. This study analyzes 221,389 user reviews from 202 women’s health apps on the Google Play Store (United States) to examine how requirement themes and affective expressions relate to user satisfaction. We employ topic modeling to identify dominant requirement themes from review text, transformer-based models to classify sentiment and discrete emotions, and multivariate regression to estimate their associations with star ratings. The analysis identifies 14 coherent requirement domains spanning core reproductive tracking, usability, technical reliability, subscription management, device integration, and community governance. Regression results show that technical reliability issues, subscription-related issues, and community governance concerns are significantly associated with lower ratings, whereas confirmation of core functionality is positively associated with satisfaction. Affective intensity further amplifies these effects, with stronger negative emotional expression linked to additional reductions in ratings. Overall, the findings highlight technical reliability, monetization transparency, and psychologically safe community spaces as central determinants of user satisfaction in women’s health digital ecosystems.
Download

Short Papers

Paper Nr:	19
Title:	A Decentralized Multichain Asset Management Platform on Blockchain Networks
Authors:	Luiz Vasconcelos Júnior, Celso Massaki Hirata and Juliana de Melo Bezerra
Abstract:	The asset management industry has traditionally relied on multiple intermediaries, leading to inefficiencies such as higher fees, slower transactions, and reduced transparency, which can hinder investor trust and capital efficiency. Decentralized Finance (DeFi) presents an alternative by leveraging blockchain technology to eliminate intermediaries, enhance transparency, and grant investors direct control over their assets. However, the DeFi ecosystem faces the significant challenge of liquidity fragmentation across different blockchains, limiting capital efficiency and complicating asset management strategies. This paper introduces a decentralized multichain asset management platform designed to address liquidity fragmentation by utilizing Lay-erZero’s omnichain interoperability protocol. The platform employs a hub and spoke architecture, with the hub chain serving as the central point for management and coordination, and spoke chains representing various blockchains where investment activities occur. By integrating LayerZero, the platform enables secure and efficient cross-chain communication, allowing managers and investors to operate across multiple blockchains from a single interface. Considering interoperability, security, and efficiency in blockchain technology, the outlined platform drives a more inclusive, resilient, and innovative asset management, serving as a reference model for future financial system developments and for emerging demands that require multichain solutions.
Download

Paper Nr:	40
Title:	Academic Dropout in Brazilian Universities: Development of Algorithms for Risk Profile Identification
Authors:	Vitória de Lourdes Carvalho Santos and Wladmir Cardoso Brandão
Abstract:	School dropout in higher education has a significant impact on students’ professional development and on the efficiency of educational institutions. This study develops and compares three predictive models (Decision Tree, Neural Network, and XGBoost) to identify dropout risk profiles among university students at PUC Minas, based on a dataset of 94,052 records enriched with institutional course information. By applying data mining and machine learning techniques, the XGBoost model achieved 94.3% accuracy and 98.0% AUC-ROC, demonstrating superior performance in identifying at-risk students. The results support proactive retention strategies and validate the importance of academic integration factors predicted by Tinto’s theoretical model.

Paper Nr:	76
Title:	A Multivocal Literature Review on Data Smells and Data Debt Nuances
Authors:	Nicolas Hahn and Afonso Sales
Abstract:	Data-driven systems increasingly depend on data pipelines, making them susceptible to subtle data quality issues. Despite growing attention to data smells and data debt, knowledge remains fragmented. This paper presents a multivocal literature review synthesizing scientific and grey literature to examine their definitions and mitigation strategies. Results show that data smells emerge from socio-technical factors (e.g., evolving requirements, poor data governance), affecting trust and maintainability. By consolidating evidence, this study provides foundation on data smells and data debt, supporting a more proactive data quality management.
Download

Paper Nr:	94
Title:	Lightweight Blockchain-Based Semantic Architecture for Chain of Custody and Traceability in LIMS Systems
Authors:	Francesco Krum, Glênio Descovi, Thiago Paulo Both, Pedro Bilar Montero, Gabriel Rodrigues da Silva, Fernando Soso Girardi, Alencar Machado and Vinícius Maran
Abstract:	Laboratory Information Management Systems (LIMS) are essential in regulated laboratory environments, where data integrity, traceability, and chain of custody directly impact auditability, legal reliability, and sanitary decision-making. However, most existing LIMS rely on centralized and mutable data repositories, which expose custody records to undetected post hoc modifications, limit interoperability across institutional boundaries, and weaken trust during audits and forensic investigations. This paper proposes a hybrid architecture that integrates Semantic Web technologies and a lightweight permissioned blockchain to address these limitations. Ontologies based on RDF/OWL are used to semantically model the collection lifecycle and enable logical consistency verification, while a Hyperledger Besu blockchain anchors cryptographic hashes of semantic records using a Proof of Authority (PoA) consensus mechanism, ensuring immutability with low computational overhead. The architecture was validated through a case study conducted on the Laboratory Portal of the Animal Health Defense Platform of Rio Grande do Sul (PDSA-RS), demonstrating that the proposed approach provides an auditable and tamper-proof history of laboratory events, improving transparency and interoperability without overburdening existing LIMS infrastructures.
Download

Paper Nr:	121
Title:	Decentralized Identity for Computing Continuum Using Blockchain
Authors:	Stefano Loss, João Pedro Barboza, Nelio Cacho, Frederico Lopes and Arthur Souza
Abstract:	In an increasingly connected world, the management of digital identities has become a critical concern, particularly with the rise of decentralized technologies. This paper explores the integration of decentralized identity management within the Computing Continuum. This paradigm leverages cloud, fog, and edge computing to address the challenges posed by the growing amount of data generated by Internet of Things (IoT) devices. By employing blockchain technology, specifically through the use of Hyperledger Identus, we present a robust framework that enhances security, privacy, and user control in distributed computing environments. It also presents details of the evolution of the SappArchi project, a middleware platform designed for smart city applications. In this way, the proposed solution introduces a new component for the SappArchi extension that supports decentralized identity via blockchain. A case study is presented in a military context to demonstrate the practical application of the proposed architecture. The evaluation of this architecture shows significant improvements in terms of security, scalability, and user autonomy. This study contributes to the ongoing research in decentralized identity management and provides a foundation for future developments in secure and scalable IoT applications.
Download

Paper Nr:	122
Title:	Trustworthy Orchestration for Edge-to-Cloud Continuum Using Blockchain
Authors:	Stefano Loss, Ian Gabriel Dias, Nelio Cacho, Frederico Lopes and Arthur Souza
Abstract:	Applications in the context of Smart Cities and the Internet of Things (IoT) deal with large volumes of data processing. For this reason, these solutions require efficient and distributed management of computational resources to maintain high-quality service metrics. The Continuum computing paradigm enhances processing efficiency through orchestration by effectively integrating computational resources across multiple levels, from the edge to the cloud. However, the orchestration of tasks among these levels must be secure and auditable, as the diverse processing nodes introduce more points of vulnerability. Therefore, blockchain technology can be the key to adding a trustworthy security layer to the orchestration process. In this work, we extend an Edge-to-Cloud Continuum architecture by introducing a new component that enhances trustworthy orchestration. This component maintains a ledger of requests with provenance data, enabling the tracking of malicious users and ensuring a secure and auditable distributed orchestration process. A real-world use case was used to test the proposed component, and the results confirm the feasibility of running a blockchain component on lower-power computational layers. This opens up opportunities for further enhancements and new functionalities involving blockchain technology within the computing continuum context.
Download

Paper Nr:	138
Title:	Hybrid Threat Detection in Real-World Academic DNS Traffic: Combining Threat Intelligence with Machine Learning
Authors:	Guilherme Romanholo Bofo, João Rafael Gregório, Leandro Alves Neves and Adriano Mauro Cansian
Abstract:	The Domain Name System (DNS) is a fundamental Internet service and a frequent target of abuse in phishing, malware distribution, spam, and command-and-control campaigns. Traditional detection approaches based on blocklists are effective for known threats, but they often fail to identify newly created malicious domains in time to prevent early exposure. This work proposes a hybrid system for malicious domain detection based on passive DNS traffic analysis that combines open-source threat intelligence feeds with a lightweight Random Forest classifier trained on lexical, active DNS, and TF-IDF features. The system was designed as a modular pipeline and deployed in a production academic recursive resolver that processes more than 20 million DNS queries per day. On test data, the best configuration achieved an AUC of 0.918 and an F1-score of 0.81. A one-year real-world deployment showed that the approach can identify 512 to 1439 malicious domains per month, while also highlighting the operational trade-off between early detection capability and false positives in highly imbalanced traffic. These findings demonstrate the system’s potential for early threat detection beyond blocklists and support the effectiveness of combining lightweight machine learning models with threat intelligence in real-world cybersecurity operations.
Download

Paper Nr:	142
Title:	Visual Requirements Modeling for Blockchain Systems: An Exploratory Software Engineering Study
Authors:	José Silva, André Araújo, Rafael Araújo and Rendrikson Soares
Abstract:	Blockchain-based systems increasingly rely on smart contracts to encode and enforce business rules in decentralized environments. Despite their growing adoption, software development in this context remains largely implementation-oriented, with limited support for systematic requirements modeling during the early stages of the software lifecycle. This paper investigates how the absence of visual requirements modeling affects the evolution and maintenance of blockchain-based software systems. An exploratory software engineering approach is adopted, combining a structured review of the state-of-the-art with an empirical analysis of five real-world blockchain projects. The results indicate that informal or incomplete requirements documentation, combined with a lack of dedicated visual modeling support, is associated with frequent, high-impact changes to smart contracts, delayed treatment of non-functional requirements, weakened traceability, and increased rework. By articulating evidence from both research and practice, this paper identifies recurring challenges and gaps and provides a conceptual and empirical basis to motivate future research on visual requirements modeling for blockchain systems.

Paper Nr:	163
Title:	How Language Concreteness Affect Crowdfunding Success: A Combination of LLM-Based Text Classification with Statistical Analysis
Authors:	Andreas Gregoriades, Michael Georgiades and Melpomeni Kasapidou
Abstract:	This study investigates the effect of linguistic concreteness on crowdfunding campaigns’ success. The research focuses on two crowdfunding categories of campaigns, the Technology and Film & Video since these two are the most popular categories and are associated with different backer motivations. Specifically, we hypothesize that technology backers tend to require detailed, functional information, whereas Film & Video backers are more driven by intrinsic motivations, thus the language used in campaign may function differently in personating these audiences. A dataset of Kickstarter campaigns is analysed using a prompt-based fine-tuned large language model (LLM) that assess the concreteness of projects’ descriptions. A logistic regression model is used with concreteness as independent variable and project success as depended variable and category of project as moderator and goal as fixed effect. The interaction of concreteness with category examines whether concreteness influences success differently by project category. The findings provide evidence that highly concrete language increases the odds of success for Technology campaigns and thus provide practical guidance for optimizing crowdfunding communication strategies.
Download

Paper Nr:	179
Title:	AranduFlow: Towards Using Performance and Domain Metadata to Predict and Optimize Dataflow Execution in HPC Environments
Authors:	André Igor Pereira, Ubiratam de Paula Junior, Daniel de Oliveira and Rafaelli Coutinho
Abstract:	Predicting execution time is a key challenge for dataflow-based applications executed on HPC environments. Variability in input data, execution parameters, and infrastructure configurations makes the search space increase exponentially, yet such estimations are important for resource-selection decisions. This paper presents AranduFlow, a dataflow-agnostic gateway that supports dataflow submission and provides execution-time predictions based on historical execution data and machine learning models. AranduFlow complements existing HPC schedulers by offering predictive decision support without interfering with scheduling policies or execution semantics. The approach is evaluated using a real-world bioinformatics dataflow executed across multiple HPC clusters under different conditions. Experimental results demonstrate that AranduFlowachieves accurate execution-time predictions, with decision tree–based models offering a favorable trade-off between predictive performance and interpretability.
Download

Paper Nr:	181
Title:	An Elastic and Auditable Privacy-as-a-Service Approach for Data Lakes
Authors:	Thiago Jordão, Wesley Ferreira, Marcos Bedo and Daniel de Oliveira
Abstract:	While Data Lakes have become a de facto standard for storing large volumes of heterogeneous data, ensuring compliance with privacy regulations such as the LGPD and GDPR remains challenging when publishing data. Compliance often requires combining multiple anonymization and privacy-preserving techniques, like k-anonymity and differential privacy, each addressing different risks. Existing solutions either do not scale well or lack the flexibility to dynamically combine techniques in a single dataflow, potentially lacking the auditability required in regulated environments. In this study, we present ARANI, a framework that provides an elastic and auditable Privacy-as-a-Service solution for large-scale data environments. ARANI enables the definition of complex anonymization dataflows through the concept of “Experiment Line”, and its orchestration and controlled execution, while ensuring the auditability of each processing step. ARANI leverages scalable and resilient execution to support concurrent, policy-diverse workflows on top of Kubernetes. We illustrate the efficacy of our approach by anonymizing real-world crime data from the Public Security Secretariat of São Paulo/Brazil, showing privacy risks can be managed without compromising utility in a scalable environment.
Download

Paper Nr:	183
Title:	PROMISE: A Process Mining Methodology for Small and Medium-Sized Enterprises (SMEs)
Authors:	Carlijn Kokkeler, Faiza Bukhsh, Lucas Meertens, Samet Kaya and Rob Bemthuis
Abstract:	Process mining provides a data-driven means to analyze, monitor, and improve business processes from event logs. Despite its recognized potential for operational insight and value creation, adoption in practice, particularly among small and medium-sized enterprises (SMEs), remains limited. Prior work offers insufficient methodological guidance that explicitly addresses SME-specific constraints. This paper presents PROMISE, a structured methodology that guides SMEs through the process mining lifecycle. PROMISE defines concrete phases, activities, and deliverables, offering a step-by-step framework for practitioners with foundational knowledge of business processes and organizational data. We develop PROMISE following Design Science Research (DSR) and evaluate it via two industrial case studies complemented by expert review. The findings indicate that PROMISE is feasible in SME settings, supports systematic execution of process mining initiatives, and can be adapted to varying organizational contexts.
Download

Paper Nr:	192
Title:	NewsSampler: News Classification from Samples Generated by Clustering
Authors:	Luíza Diapp, Lisiane Reips, Aurora T. R. Pozo and Carmem S. Hara
Abstract:	The increasing volume of online news requires scalable methods for filtering and organizing information with minimal manual effort. This paper presents NewsSampler, a workflow that combines clustering-based sampling with large language models to improve relevance classification in large, unlabeled news collections. The approach first groups documents into thematic clusters and selects representative samples, which are manually labeled and used as few-shot examples to classify the remaining data. Two text representations (TF-IDF and BERT) and two clustering algorithms (K-Means and DBSCAN) were evaluated on two datasets related to the "Iguaçu River" and the "Portuguese man-of-war". Results indicate that TF-IDF combined with K-Means yields the most coherent thematic segmentation. Using the selected examples, GPT achieved 71.9% precision and 93.6% accuracy, outperforming a clustering-based baseline. These findings demonstrate that clustering-guided representative sampling can reduce labeling effort while enabling effective news relevance classification.
Download

Paper Nr:	199
Title:	A Practical Approach to Generating and Exploring Large-Scale Datasets for Face Recognition
Authors:	Jean Magalhães, Raylander Marques Melo, José Maria Monteiro, José Wellington Franco, Javam Machado and Cesar Barreira
Abstract:	Currently, facial recognition is an essential task across a wide range of applications, including user authentication, person identification, and image organization in social networks. The popularity of facial recognition technology is primarily attributed to its simplicity, efficiency, and practicality. These characteristics make facial recognition an especially advantageous alternative for use in digital government systems and public security. However, in these scenarios, an individual’s image is compared against a database containing thousands of faces, which poses significant challenges for the development of such applications. In this paper, we present a practical approach to generating and utilizing large-scale datasets for facial recognition. Initially, we introduce a strategy for building a broad, diverse collection of facial images. Subsequently, using this technique, we constructed a massive facial image dataset, named Faces4LA, which is publicly available. The Faces4LA dataset includes both real and synthetic images, along with provenance information. Finally, we propose a practical guide to support the development of large-scale facial recognition applications and conduct a case study to evaluate the proposed approach.
Download

Paper Nr:	201
Title:	Unleashing the Value of Metadata: An Extensible Architecture for Holistic Metadata Management in Enterprises
Authors:	Jan Schneider, Christoph Gröger, Arnold Lutsch, Holger Schwarz and Bernhard Mitschang
Abstract:	In the digital age, the analysis of data is crucial for enterprises of almost all industries. To support the collection, management, and analysis of data, modern data platforms such as data lakehouses have become prevalent in recent years. Despite these technological advances, metadata management in today’s enterprise data architectures still shows serious shortcomings, as metadata is only inadequately captured and locked up in silos. As a result, the knowledge transported by this metadata cannot be fully exploited, which complicates and slows down all metadata-based applications. In this paper, we first investigate common technical challenges of metadata management in enterprise data architectures. Secondly, we derive requirements that enterprise data architectures must meet to address these challenges. By considering these requirements, we then present a conceptual architecture as solution, which utilizes a knowledge graph and a microservice architecture to enable holistic management of metadata and to provide it to different applications in a flexible manner. Finally, we evaluate and discuss this architecture with the help of an early prototype.
Download

Paper Nr:	215
Title:	A Proposal for an LLM-Based Multi-Agent Architecture for Orchestrating Access to Multiple Data Sources
Authors:	João Pedro V. Pinheiro, Yenier T. Izquierdo, Luiz André P. P. Leme, Antonio L. Furtado and Marco A. Casanova
Abstract:	This paper addresses the problem of accessing multiple data sources that possibly store heterogeneous multimodal data. It proposes an LLM-based multi-agent architecture that breaks complex data-access tasks into smaller ones and uses well-known agent communication protocols, along with a Pub/Sub broker, to orchestrate the agents. For easy access, it offers a natural language conversational interface that allows users to formulate their questions in natural language and to interact with the interface until their request is understood. The paper also includes an end-to-end example that illustrates how the architecture processes a user request, using multimodal data obtained from social media.
Download

Paper Nr:	221
Title:	The Limits of Behavioral Pruning: Why Enterprise RAG Lifecycle Management Is a Governance Problem
Authors:	Aaditya Chauhan
Abstract:	Enterprise RAG systems accumulate embedding corpora that grow indefinitely, yet lack principled lifecycle management for their vector indexes. We propose behavioral pruning—using access patterns and semantic redundancy to select which documents to retain in the active index—centered on a governance mechanism we call semantic anchors: documents with no near-duplicates above a similarity threshold are preserved in the active index regardless of access frequency, and near-duplicate clusters retain their newest member, providing a traceable justification for retention decisions. We present an empirical evaluation identifying both conditions where the framework succeeds and where it fails. The recall advantage of behavioral pruning is conditional on corpus characteristics: random pruning achieves higher recall on historical queries (0.118 ± 0.011) than all structured methods; density-weighted behavioral pruning improves recall on fresh queries by 12% over semantic-only pruning. A supplementary simulation on BEIR benchmarks confirms behavioral signals help on redundant corpora with correlated access; on diverse corpora with uncorrelated access, recall degrades by up to 31%.
Download

Paper Nr:	226
Title:	Impact and Applications of Aggregation Methods in Fuzzy Rule-Based Classification Systems: A Systematic Review
Authors:	João Pedro Stone Moreira, Giancarlo Lucca, Lizandro S. Oliveira, Diego Duarte Bottero, Rafael Berri and Bruno Dalmazo
Abstract:	The inherent uncertainty of real-world data justifies the use of fuzzy logic; however, the effectiveness of these systems is intrinsically linked to how rules are aggregated during the reasoning process. This study presents a Systematic Literature Review (SLR) focused on aggregation methods in Fuzzy Rule-Based Classification Systems (FRCS). The study was based on the Kitchenham & Charters guidelines. After defining the research questions, search combinations were entered into the IEEE Xplore, Scopus, ACM Digital Library, and Web of Science digital libraries. After obtaining partial results from 62 studies, exclusion criteria were applied to filter the most relevant studies, leaving 10 studies for analysis. The results revealed a shift towards non-linear aggregation operators, such as generalizations of the Choquet Integral and non-associative overlap functions. This study concluded that these methods improve the flexibility and generalization capacity of the model, capturing complex interactions between the criteria. The findings also indicated that the proposed aggregation mechanisms significantly improve classification accuracy across various domains, including network attack detection and image processing, while also providing better control over the propagation of uncertainty.
Download

Paper Nr:	238
Title:	Toward a Time Series-Specific Machine Learning Life Cycle: Challenges and Requirements
Authors:	Christoph Schrade, Theo Zschörnig and Bogdan Franczyk
Abstract:	Machine learning (ML) life cycles differ from conventional software development life cycles because model behavior depends not only on code but also on data. However, the form of an ML life cycle also depends on the data modality underlying the application. Focusing on time series applications, we argue that time series data exhibits characteristics that make its use in ML particularly challenging. We compare key data handling activities, especially data collection and data processing, in time-series-based ML applications with common practices for text and image data. On this basis, we identify central challenges associated with time series and derive requirements for a time series-specific ML life cycle. The paper therefore provides a first step toward a future reference model for developing and operating time series-based ML systems.
Download

Paper Nr:	274
Title:	Framework for Assessing Analytics and Artificial Intelligence in Portfolio Project Management
Authors:	Héctor Melo, María del Pilar Villamil and Oscar Ávila
Abstract:	The increasing integration of Analytics and Artificial Intelligence (AI) into Project Portfolio Management (PPM) has generated significant academic and industrial interest; however, a structured and operational framework to systematically assess their contribution across PPM domains remains absent. This study proposes an evaluation framework that integrates foundational PPM standards (ISO 21504, PMI Standard for Portfolio Management, and AXELOS MoP) with established Analytics and AI conceptual models derived from academic literature, and industry methodologies such as CRISP-DM and ASUM-DM. The framework is structured around categories, criteria, and evaluation questions, enabling the systematic assessment of analytical approaches, learning paradigms, models, evaluation metrics, lifecycle phases, and business impact across portfolio practices. The framework is applied through a structured literature review, demonstrating its analytical capacity to classify and compare heterogeneous research contributions. The proposed framework provides a replicable instrument to support consistent academic analysis and to identify research gaps in the application of Analytics and AI within Project Portfolio Management.
Download

Paper Nr:	308
Title:	An Incremental E-Commerce Review Dataset for Sentiment Analysis in Portuguese Language
Authors:	Diego Duarte Bottero, Joelson S. Junior, João P. S. Moreira, Lizandro de S. Oliveira, Giancarlo Lucca, Bruno L. Dalmazo, Rafael A. Berri and Eduardo N. Borges
Abstract:	Datasets are essential for advances in artificial intelligence, particularly in natural language processing tasks such as sentiment analysis. However, the scarcity of large, updated datasets in Brazilian Portuguese remains a challenge. This study presents the development of an automated web scraping pipeline to collect product reviews from e-commerce platforms. A pre-processing pipeline was applied to standardize and clean the textual data, improving its quality and usability. The results show that incremental data collection captures important lexical shifts across rating classes, better reflecting current consumer behavior. Additionally, to promote Open Science, the newly created dataset will be publicly released to support transparency, reproducibility, and collaborative research. Future work includes expanding the scraping process to new domains and applying machine learning techniques to this dataset.
Download

Paper Nr:	322
Title:	Active Rules for Embedded Databases: Lightweight Event-Driven Query Processing for Resource-Constrained Devices
Authors:	MacKenzie Richards and Ramon Lawrence
Abstract:	Resource-constrained Internet of Things (IoT) and edge devices increasingly require autonomous, real-time decision making without relying on cloud services or energy-intensive polling loops. This paper presents a lightweight active rule mechanism for event-driven query processing in an embedded database environment. The design adds event-condition-action style rules that are evaluated on insertion, supports sliding-window aggregates and user-defined callbacks, and reuses the existing query framework to provide active database functionality with low overhead. Existing active database and stream processing systems support event-condition-action (ECA) rules, but they target server or embedded Linux environments and are not designed for microcontroller-class devices with only a few kilobytes of memory. Experimental results show that active rules are practical on embedded hardware. On an Arduino Due, a representative frost-detection rule using a 10-record sliding window sustained about 1909 inserts per second with an average overhead of 185 µs per insert, and performance scaled predictably with window size. Comparison with manually issuing equivalent user-level queries after each insert showed nearly identical performance, indicating that the measured cost is dominated by query execution rather than the active rule interface itself. On a computer, the active rule implementation processed a 10-record window query in 0.00139 ms per insert, compared to 1.897 ms per insert for InfluxDB with Kapacitor, highlighting the advantage of local embedded rule execution over a client-server stream processing architecture. These results demonstrate that active database functionality can be implemented efficiently on highly constrained devices and can support low-latency local monitoring, alerting, and actuation for edge and IoT workloads.
Download

Paper Nr:	326
Title:	Contrastive Learning for Cross-Modal Entity Matching: An Experimental Study
Authors:	Paulo Henrique Santos Lima and Leonardo Andrade Ribeiro
Abstract:	Cross-modal Entity Matching aims to determine whether two representations from different modalities, such as a textual description and an image, refer to the same real-world entity. In this paper, we present an experimental study on the use of CLIP for cross-modal text-image EM, evaluating three architectures: (i) a zero-shot model based on cosine similarity with temperature scaling; (ii) CLIP-FT, a fine-tuned variant; and (iii) CLIP-MLP, a fine-tuned variant with a two-layer MLP head. All fine-tuned models are trained with different combinations of loss functions. Experiments on five publicly available datasets show that tuning the temperature parameter is essential for the zero-shot baseline, improving F1 by up to 29%, and that combining the original CLIP loss with a binary classification loss improves matching performance. CLIP-FT outperforms both the baseline and CLIP-MLP on all five datasets, achieving up to 0.969 F1, while also being 32–53% faster to train than CLIP-MLP, establishing itself as the most accurate and efficient architecture for cross-modal EM in our study.
Download

Paper Nr:	20
Title:	Towards a Blockchain and Social Media-Based Architecture for Project Fundraising
Authors:	Bryan Diniz Borck, Celso Massaki Hirata and Juliana de Melo Bezerra
Abstract:	Social media platforms enable users to reach a wide audience quickly and effectively. By engaging directly with followers, users can foster a sense of community and trust, which are crucial for the success of crowdfunding. In parallel, blockchain has gained prominence as a reliable and transparent infrastructure for conducting financial transactions. This paper presents an innovative solution to fundraising for artists’ projects by combining blockchain technology with social influence to create a decentralized and secure system. We propose an architecture that allows artists to raise funds directly from their followers, using the blockchain-based social media Farcaster and its decentralized applications called Frames. The main idea is to eliminate dependence on traditional financial intermediaries and democratize access to funding, providing decentralized financial transactions that are direct, verifiable, and secure. By focusing on practical implementation and emerging technologies, our work represents a significant advance in how resources are raised and managed, offering a valuable reference for future decentralized funding systems.
Download

Paper Nr:	35
Title:	A Blockchain-Based Pharmaceutical Traceability System with Neural Anomaly Detection Using Hyperledger Fabric
Authors:	Julio E. Asillo Mendoza, Edwin A. Lopez Huaman and Rosa Felix
Abstract:	The counterfeiting of pharmaceutical products remains a critical public health threat, with an estimated 10% of medicines in developing countries classified as substandard or falsified. These products generate severe consequences, including therapeutic failure, toxicity, antimicrobial resistance, and significant economic losses. This study proposes a decentralized web-based system that integrates a Hyperledger Fabric blockchain with deep neural network models to ensure end-to-end traceability and authenticity verification of pharmaceutical products. The system records supply-chain events through smart contracts on an immutable ledger and performs real-time anomaly detection over lifecycle data. A pilot deployment with real pharmaceutical batches was conducted to validate performance and operational feasibility. Results show a 100% verified traceability rate from laboratory to pharmacy and 93% from laboratory to end-user. The system sustained over 99.9% transaction success with confirmation latencies below 5 seconds, demonstrating a scalable, transparent, and high-performance architecture capable of mitigating adulteration risks in pharmaceutical supply chains.

Paper Nr:	50
Title:	From Commits to Code Smells: A Developer-Centric Visualization of Technical Debt
Authors:	Manoel Valerio da Silveira Neto, Andreia Malucelli and Sheila Reinehr
Abstract:	Source code quality is crucial to overall software quality and is influenced by the developer’s experience and practices. However, developers often submit code without being aware of its quality, and code smells are typically detected only after the fact through static analysis tools. This reactive approach frequently leads to rework and an increase in technical debt (TD). This paper proposes a developer-oriented process for monitoring and managing TD caused by code smells. By combining process mapping with the mining of source code repository data, the proposed approach aims to provide developers with insights into their code smell generation trends, thereby supporting improvements in coding practices and the reduction of TD introduction. The approach is evaluated through experimental software engineering and knowledge discovery in databases. The results demonstrate its ability to generate actionable knowledge about the code smell creations, as well as to identify behavioral and temporal patterns that support the continuous improvement of software development practices.
Download

Paper Nr:	139
Title:	Predictive Modeling for Low-Voltage Network Dimensioning in Energy Regularization Areas Using Neural Networks
Authors:	Patrick C. Araujo, Pedro V. Bernhard, Thiago P. Freire, Joana K. A. Silva, Eduardo F. P. Dutra, Levi C. Santos, Weslley K. R. Figueredo, Italo F. S. Silva, João D. S. Almeida, Luis J. E. R. Cabrejos, Aristofanes C. Silva, Auriane A. M. Santos, Felipe M. Feyh, Carlos J. S. Moura, Matheus Menezes and Lucas P. A. Pinheiro
Abstract:	This work proposes a predictive pipeline to estimate the length in meters of low-voltage networks in energy regularization areas based on the position of meters. In a dataset with 3.5 million meters and 20,526 network segments from a Brazilian distributor, we grouped meters via DBSCAN, extracted geometric and structural attributes (cluster polygon and MST), and trained regressions (MLP, XGBoost, Random Forest, and ChebyKAN). MLP performed best, with MAE ≈ 107m, RMSE ≈ 154m, and R 2 ≈ 0.91 in cross-validation, and low variability between folds. The contribution is a reproducible workflow, integrated with GIS, that anticipates project inputs and purchases in contexts with limited cartography. We discuss hyperparameter choices and practical implications for utilities, as well as limitations and future directions (other regions and clustering alternatives).
Download

Paper Nr:	157
Title:	Energyguard Data Lake Architecture: From Heterogeneous Ingestion to Federated Data Sharing
Authors:	José L. Hernández, Vagelis Karakolis, Theodosios Pountridis, Spiros Mouzakitis, Huy Nam Nguyen and Gabriele Balzano
Abstract:	The increasing deployment of artificial intelligence, digital twins and advanced sensing technologies in the energy sector requires data infrastructures capable of managing heterogeneous sources, large-scale time series and strict security, privacy and governance requirements. This paper presents a modular Data Lake architecture designed to support trustworthy AI development, testing and validation in energy-related scenarios. The proposed solution provides a scalable pipeline for data ingestion, harmonisation and storage, while promoting interoperability through semantic modelling, metadata management and standardised interfaces. A layered architecture enables integration with external services and federated data sharing, ensuring controlled access, traceability and accountability across the data lifecycle. The resulting design provides a practical blueprint for secure and interoperable Data Lake platforms supporting AI-enabled energy systems experimentation.
Download

Paper Nr:	161
Title:	Approaching Network Traffic Classification by Exploratory Data Analysis Insights Extracted from the StreamDataNetClass
Authors:	Gabriel R. O. Silva, Bruna Novack, Giancarlo Lucca, Renata Reiser, Adenauer Yamin, Helida Santos, Lizandro S. Oliveira, Bruno M. P. Moura and Eduardo M. Monks
Abstract:	Artificial intelligence has attracted considerable attention in recent years; however, its effective application depends on the availability of high-quality, well-curated datasets for training robust models. In this context, accurate data targeting and classification are essential. Network traffic classification, especially the streaming traffic data, has received increasing attention from both the technological and academic communities due to the continuous growth in the number of connected devices and users. This study presents an Exploratory Data Analysis of a new promising dataset created explicitly for network streaming traffic classification, named StreamDataNetClass. The analysis employs a variety of visualization techniques, including bar plots, box plots, scatter plots, and kernel density estimation graphs, to characterize the dataset’s behavior and underlying patterns. At the end, beyond the discussion of the observed results, we understand that the StreamDataNet-Class has robust data, which can inform different aspects of this type of network flow.
Download

Paper Nr:	220
Title:	Master Data Management in Practice: An Empirical Analysis of Architectural and Governance Capability
Authors:	Gabriel Pereira, André Araújo, José Silva and Rodolfo Cavalcante
Abstract:	Master Data Management (MDM) is widely recognized as a foundational component of data governance, supporting data quality, interoperability, and cross-system consistency. Although the literature proposes architectural models, governance frameworks, maturity assessments, and AI-enabled data quality mechanisms, there is limited empirical evidence on how these concepts are operationalized in real-world software projects. This study investigates the alignment between the state of the art and the state of the practice through a structured empirical evaluation. A literature review identified key analytical dimensions of MDM, which were used to design an evaluation instrument applied to eight software projects. The results reveal heterogeneous adoption levels, fragmented architectural capabilities, partial formalization of governance, and inconsistent implementation of lifecycle and entity-resolution mechanisms. While isolated MDM practices are frequently observed, their integration into coherent institutionalized strategies remains limited. The findings highlight a structural gap between conceptual models and operational implementation, emphasizing the need for empirically grounded frameworks to assess architectural coherence, governance integration, and sustained master data reliability.

Paper Nr:	252
Title:	Named Entity Recognition for Sensitive Data Recognition in the Philippine Financial Industry Context
Authors:	Raymond Lin Chen and Charibeth Cheng
Abstract:	With the Philippines' fintech industry rapidly growing, accurate detection of sensitive data is crucial for regulatory compliance under the Data Privacy Act of 2012. Named Entity Recognition (NER) is commonly used for sensitive data identification but lacks customization for local financial contexts. This study proposes a localized NER methodology integrating structured KYC data, comparing LSTM and LLM-based approaches. The LSTM model achieved the highest recall (95.1%) and fastest runtime (0.0958 seconds) for binary classification, ideal for real-time applications. However, multiclass recall dropped to 55.5%, reflecting increased complexity. Zero Shot, Five Shot, and RAG approaches showed greater interpretability but slower runtimes and higher costs. Cost analysis favors LSTM for large-scale deployments. This research provides guidance on sensitive data classification and regulatory compliance, emphasizing the need for ongoing model iteration tailored to Philippine fintech challenges.
Download

Paper Nr:	289
Title:	Ensuring Data Integrity on ETL Systems Using Blockchain
Authors:	Rodrigo Rodrigues, Bruno Oliveira and Orlando Belo
Abstract:	Data integrity is crucial for supporting decision-making processes. Extract, transform, and load systems are key components for ensuring data integrity and handling data extraction, cleaning, optimization, and loading into data warehouses. As demands for security, integrity, and traceability grow, blockchain technology arises as a transformative solution. This paper proposes a strategy for integrating blockchain into ETL processes, addressing system architecture, data integrity mechanisms, and the interactions among system components. Two use cases are presented. The first one was selected for ensuring the integrity of fact tables through a hash-based registration system that enables data verification during system updates, and the second for ensuring the integrity of slowly changing dimensions (type 4) tables using an immutable change-log mechanism implemented via blockchain. A prototype system was developed and deployed in a specific data warehousing populating system scenario, validating the proposed strategy and demonstrating the feasibility and benefits of blockchain-based data integrity assurance in data warehousing environments.
Download

Area 2 - Artificial Intelligence and Decision Support Systems

Full Papers

Paper Nr:	25
Title:	Digital Image Processing to Enhance Strain Analysis on Grid Marked Stamped Steel Pieces
Authors:	Guilherme Razzini Boeger, Victor Hugo Moresco, Vinicius Thiago Lecco Rampinelli, Roan Sampaio de Souza, Jetson Ferreira Lemos, Marden Valente de Souza, Francislaynne Lages Dias and Rafael Stubs Parpinelli
Abstract:	Stamping is a standard process in the Automotive Industry where a die presses steel sheets into complex parts, like car doors. However, because undesirable phenomena can occur during stamping, a post-stamping analysis is essential. This is the process of strain analysis, aiming to ensure the performance and security of the final product. A method of strain analysis is through grid marking (printing a grid pattern on the blank steel sheet). Once stamping finishes, the grid will have deformed, and a commercial image-based strain analysis software is employed to analyze it. Based on pictures of grid-marked stamped steel pieces, the software identifies grid crossing points and reconstructs them in a digital environment. However, the pictures can lack quality (e.g., illumination), negatively impacting the digital reconstruction. Therefore, this work proposes pre-processing the images using deterministic digital image processing alongside Deep Learning-based Super-Resolution to improve the digital reconstruction of the piece, broadening the depth of strain analysis. Results showed crossings identification increases up to 260%, and statistical tests confirm that the recommended approach significantly outperforms the remaining methods, showing no statistical equivalence to the other analyzed approaches at a significance level of α = 0.05.
Download

Paper Nr:	28
Title:	An Iterated Local Search-Based Algorithm for the Short-Term Technology Maintenance Scheduling: A Case Study in a Bulk Solid Port
Authors:	Breno Augusto Klein Magnago and Luciano Perdigão Cota
Abstract:	Technology asset maintenance became relevant in maintenance planning and control due to the technological modernization of industrial plants. This study introduces the short-term technology maintenance jobs scheduling problem. In this problem, a set of maintenance jobs needs to be planned for teams with different skill sets within a given time window. Each team has a certain number of employees available to perform the jobs. Additionally, each job is associated with a specific technology asset, accompanied by prioritization criteria and designated time windows for execution. The objective is to minimize the number of teams and displacements, as well as the penalties for not executing jobs, taking into account the asset’s priority and type. This problem introduces new characteristics into the literature, including constraints on improvement and corrective jobs, asset prioritization, individual employee control within teams, and their displacement in industrial plants. Because this problem variant is NP-hard, only heuristics can solve large instances within a reasonable amount of computing time for decision-making. We proposed a specialized algorithm based on Iterated Local Search to solve the problem, incorporating a heuristic to generate a good initial solution, a local search to refine it, and a perturbation procedure. The proposed algorithm was validated using a case study from a bulk solid port, and results were compared with those of another algorithm from the literature that won a recent challenge for a maintenance scheduling problem. The proposed algorithm produced better results in 85% of the instances, demonstrating its potential to support decision-making.
Download

Paper Nr:	46
Title:	LLM-Assisted Natural Language Configuration of Complex Industrial Products
Authors:	Gabriel Gobbo Araújo Valera, Marcos Alves dos Santos and Alison R. Panisson
Abstract:	Configuring complex industrial products is a demanding task, often overwhelming non-expert users due to the large number of interdependent parameters and constraints involved. This work presents an approach that integrates Large Language Models (LLMs) into the product configuration process, enabling requirements to be specified using natural language. Free-text user inputs are translated into structured configuration parameters, which are then processed by a constraint-based solver to generate valid product configurations. A case study on synchronous alternators demonstrates that the LLM-enhanced configurator accurately interprets user requests and produces solutions that satisfy all technical constraints. The results indicate improved usability for both engineers and customers while preserving high configuration accuracy. Overall, this study advances intelligent software engineering by bridging natural language interfaces and configuration solvers, providing a copilot-like system that streamlines the configuration of complex industrial products.
Download

Paper Nr:	68
Title:	Grinding Ball Replacement Planning Problem: An Improvement to the State-of-the-Art Algorithm
Authors:	Daniel Luiz de Souza, Luciano Perdigão Cota and Marcone Jamilson Freitas Souza
Abstract:	This work addresses the grinding ball replacement planning problem (GBRPP). The objective is to schedule grinding ball replacement in ball mills to maintain the particle size at the mill output within recommended values. In addition to determining the start time of each replacement, the problem also involves determining the bulk weight of grinding balls to be replaced at each instant of the planning horizon. In this work, we present an improvement to the Enhanced Iterated Local Search (E-ILS) algorithm, the state-of-the-art method for GBRPP. This improvement consists of adaptively generating a higher-quality initial solution based on the difference between the recommended and estimated power for each mill at each planning instant. Computational results obtained on real GBRPP instances showed that this procedure accelerates the attainment of high-quality solutions, reducing processing time by 26% and the average deviation from the best-known solutions from 1.02% to 0.53%. The E-ILS, incorporating this new procedure, was embedded into a decision-making system at a mining company.
Download

Paper Nr:	77
Title:	ProMoAgentAI: A Multi-Agent Collaborative Framework for Robust BPMN Generation with Cumulative Learning
Authors:	Mahmoud B. M. AbdelWahab, Sherif A. Mazen and Iman M. A. Helal
Abstract:	Business process model and notation (BPMN) is one of the most commonly used standards in process model representation and formalization. There are tools that automate business process modeling using Large Language Models (LLMs). However, these tools suffer from significant limitation whenever they retry model generation, as they repeat identical mistakes. In other words, they are not able to learn from the failures occurred in the same working session. In this research, we propose ProMoAgentAI, which is a multi-agent framework that implement a persistent session memory combined with collaboration of specialized agents. The developed system orchestrates eight distinct agents via CrewAI platform. Our key contribution is the session-memory mechanism which maintains records of all attempted solutions, encountered errors, and partially successful results. This allows agents to prevent repeating failed approaches and instead build upon the working parts from previous tries. The proposed framework implements a three-tier orchestration strategy having automatic fallback capability, which can adapt according to varying levels of complexity of processes. Through rigorous experimental evaluation on 20 PET dataset processes, we demonstrate that ProMoAgen-tAI achieves 24–28% quality improvement over the ProMoAI baseline, with 100% valid BPMN generation. Crucially, our ablation study reveals that the multi-agent architecture is essential rather than merely beneficial: without specialized agents, error recovery rates drop by 27% and valid generation rates decline from 100% to 75–81%. Industrial deployment on 540 telecom processes achieved 83% correctness, demonstrating production-ready performance. This research contributes toward development of more resilient AI-based systems capable of learning behavior for process modeling automation.
Download

Paper Nr:	81
Title:	ECHO-PM: A Hybrid NLP Decision-Support Pipeline for Emotion-Aware, NVC-Guided Message Rewriting in Distributed Software Projects
Authors:	Fellipe Anchieta Silva Barros, David Ferreira Brandão, Cleyton Mário de Oliveira Rodrigues and Aêda Monalliza Cunha de Sousa
Abstract:	Socio-emotional friction in short technical messages (e.g., frustration or dismissive phrasing) can amplify misunderstandings in distributed software development. This paper presents ECHO-PM, a hybrid NLP decision-support pipeline that (i) labels sentiment polarity, (ii) infers fine-grained emotions, (iii) derives tone-oriented signals, and (iv) optionally proposes rewrites guided by a transparent heuristic policy inspired by Nonviolent Communication (NVC), while keeping authors in control of the final message. We demonstrate the pipeline on 677 GitHub issue/discussion messages selected after filtering 1,046 initial records. Neutral sentiment is the most frequent class (56%), followed by negative (35%) and positive (9%), and emotion–sentiment heatmaps show that negative messages are not homogeneous under polarity alone. Because no manually validated tone ground truth is available at scale, tone signals are examined via agreement patterns across automatic approaches and a binary presence view (tone detected vs. not detected), avoiding predictive performance claims. Rewriting outputs are described using ROUGE-L overlap and log-perplexity as proxies, complemented by qualitative examples and an explicit discussion of validity threats. Overall, ECHO-PM provides a transparent, caution-oriented workflow for surfacing affective cues and suggesting de-escalation-oriented reformulations, without claiming controlled user-validated effectiveness.

Paper Nr:	91
Title:	Benchmarking Small Language Models for Medical Question Answering in Resource-Constrained Environments
Authors:	Marcelo Moreira West, José Amancio Macedo Santos, Fernando Brito Abreu, Manoel Gomes de Mendonça Neto and Glauco de Figueiredo Carneiro
Abstract:	While Large Language Models (LLMs) demonstrate significant potential in identifying complex medical patterns, their deployment in healthcare is restricted by critical barriers: the misalignment between standardized benchmarks and real-world patient consultations, data privacy risks inherent in commercial APIs, and the prohibitive computational costs of local execution. This study investigates the feasibility of Small Language Models (SLMs)-defined here as architectures with fewer than 10 billion parameters -serving as a proxy for high-end edge computing devices, as a viable solution to these constraints. We conduct a statistically comparative evaluation of domain-specific and general-purpose models from the Hugging Face ecosystem. By analyzing both their clinical effectiveness (comprehensiveness and hallucination rates) and computational efficiency on the K-QA dataset, we assess their viability in privacy-conscious, resource-constrained settings. Our results reveal a distinct trade-off between information density and factual reliability: while general-purpose models achieved higher comprehensiveness at the cost of increased hallucinations, clinically conservative models minimized errors but lacked coverage. Ultimately, we identify that balanced architectures, such as MedGemma-1.5-4b, offer the most promising path for deploying safe, accessible, and efficient AI decision support in resource-constrained medical environments.
Download

Paper Nr:	100
Title:	Privacy Risks of Generative AI in Organizations and Governance Models for Risk Mitigation
Authors:	Stefano Luppi Spósito, Muhtasim Azad, Cristina Fontes, Fábio Lúcio Lopes de Mendonça and Edna Dias Canedo
Abstract:	Context: The adoption of Generative Artificial Intelligence (GenAI), especially Large Language Models (LLMs), is accelerating across organizational processes that handle personal, sensitive, and regulated data. Alongside productivity gains, recent evidence has raised concerns about new and evolving privacy risks that challenge traditional data protection and compliance practices. Goal: This paper aims to synthesize the main privacy risks associated with GenAI in organizational deployments and to identify governance-oriented mechanisms that can mitigate these risks beyond isolated technical controls, with particular attention to high-impact public-sector settings. Method: We conducted a structured synthesis of recent literature on privacy risks and attacks in LLM-based systems, privacy and risk assessment methodologies, and AI governance frameworks. The evidence was organized into recurrent privacy risk categories and corresponding governance mechanisms, resulting in (i) a summary of representative studies and risks and (ii) a mapping between privacy risk categories and governance-oriented mitigation strategies. Results: The synthesis indicates that privacy risks in GenAI are dynamic, context-dependent, and socio-technical, emerging from the interaction between models, data, users, and organizational practices. Key risk patterns include data memorization and inference-based leakage, prompt-driven disclosure and exfiltration, and lifecycle emergent risks amplified by shadow AI and fragmented accountability. Effective mitigation therefore requires integrated governance models combining technical safeguards with lifecycle-based risk management, continuous privacy assessment, clear accountability structures, and auditable regulatory alignment. We also derive practical recommendations for public-sector governance, emphasizing institutional oversight, continuous monitoring, and inter-organizational data governance to sustain privacy protection and public trust.
Download

Paper Nr:	101
Title:	Interpretable Markov-Based Spatiotemporal Risk Surfaces for Missing-Child Search Planning with Reinforcement Learning and LLM-Based Quality Assurance
Authors:	Joshua Castillo and Ravi Mukkamala
Abstract:	The first 72 hours of a missing-child investigation are critical for successful recovery. However, law enforcement agencies often face fragmented, unstructured data and a lack of dynamic, geospatial predictive tools. Our system, Guardian, provides an end-to-end decision-support for missing-child investigation and early search planning. It converts heterogeneous, unstructured case documents into a schema-aligned spatiotemporal representation, enriches cases with geocoding and transportation context, and provides probabilistic search products spanning 0–72 hours. In this paper, we present an overview of Guardian as well as a detailed description of a three-layered predictive component of the system. The first layer is a Markov chain, a sparse, interpretable model with transitions incorporating road accessibility costs, seclusion preferences, and corridor bias with separate day/night parameterizations. The Markov chain’s output prediction distributions are then transformed into operationally useful search plans by the second layer’s reinforcement learning. Finally, the third layer’s LLM performs post hoc validation of layer 2’s search plans prior to their release. Using a synthetic but realistic case study, we report quantitative outputs across 24/48/72-hour horizons and analyze sensitivity, failure modes, and tradeoffs. Results show that the proposed predictive system with the 3-layered architecture, produces interpretable priors for zone optimization and human review.
Download

Paper Nr:	104
Title:	Towards a Framework Using RAG and Graph Networks in the Detection of Anomalies in Livestock Movement
Authors:	Gabriel Rodrigues da Silva, Gustavo Scopel Moz, Gabriel Vieira Casanova, Francesco Krum, Glênio Descovi de Freitas, Brunele Weber Chaves, Alencar Machado and Vinícius Maran
Abstract:	The intense volume of livestock movements poses significant challenges for sanitary and fiscal auditing, creating a complex network where fraudulent patterns, such as cattle laundering, easily blend into legitimate trade. This paper introduces a feasibility study for a novel neuro-symbolic framework that integrates Temporal Graph Networks (specifically Graph Attention Networks - GAT) with Retrieval-Augmented Generation (RAG). Unlike traditional tabular methods, our framework models the supply chain as a dynamic graph to detect structural anomalies. To validate this approach, we conducted comparative experiments using real-world data from Rio Grande do Sul. Results demonstrate that while a baseline approach using direct Large Language Model (LLM) analysis on raw data suffers from frequent hallucinations and fails to capture temporal loops, the proposed GAT model successfully isolates non-trivial anomalies. Furthermore, we outline how the RAG component converts these opaque technical alerts into legally grounded, human-interpretable reports. This work establishes the architectural viability of combining geometric deep learning with generative AI to enhance regulatory oversight while adhering to Responsible AI principles. This framework provides a foundation for future scaling in cloud-native environments and the full implementation of the RAG pipeline to ensure human-in-the-loop oversight and regulatory integrity in real-world auditing scenarios.
Download

Paper Nr:	106
Title:	An Exploratory Study of the Use of AI for Report Generation and Support for Criminal Classification in Police Incidents
Authors:	Alexandre Barbosa Santana and Glauco de Figueiredo Carneiro
Abstract:	The drafting of technical police incident reports is an essential yet resource-intensive task within law enforcement. This study investigates the utility of Large Language Models (LLMs) in supporting this workflow, first by generating a synthetic dataset of structured narratives using the gpt-4o-mini architecture. Subsequently, we propose and execute a quantitative validation methodology where a panel of simulated experts (Llama-3-8B, Gemma-2-9B, and GPT-OSS-120B) evaluates the reports from the perspective of multiple professional “personas.” The analysis identifies a central “Agreement Paradox”: while categorical agreement (Weighted Kappa) remains low to moderate, the proximity of evaluations (Mean Absolute Error) is consistently high. These findings demonstrate that despite distinct architectural scoring biases-ranging from the “benevolence bias” of smaller models to the “analytical granularity” of larger ones-the evaluators share a functionally convergent directional understanding of report quality. This study establishes the technical feasibility of AI-led auditing in specialized legal domains, emphasizing that model scale is a critical determinant of evaluative reliability.
Download

Paper Nr:	110
Title:	While My Gradient Gently Vanishes: Exploring Categorical Cross-Entropy Pitfalls in Classification Problems
Authors:	Mateus Coelho Silva and Ricardo Augusto Rabelo Oliveira
Abstract:	Classification problems are a fundamental part of Machine Learning. Prediction models are often trained using Supervised Learning, where the model is adjusted by minimizing an error function, known as the Loss Function. Deep neural networks use this strategy to optimize their parameters and minimize this loss. The most used loss function in classification is the Categorical Cross-entropy. Nevertheless, the performance of machine learning models degrades with noisy labels. Therefore, in this work, we explore the performance degradation of deep neural networks using different loss functions, namely Categorical Cross-Entropy (CCE) and Mean Squared Error (MSE). Our results show that deep neural networks trained with MSE as the loss function are more robust to noisy labels than those trained with CCE, especially at noise levels below 20%. These results indicate that the CCE function might not be the best option in all cases, especially when dealing with noisy-labelled datasets.
Download

Paper Nr:	127
Title:	When to Rerank? Adaptive Reranking Decisions for Efficient Information Retrieval
Authors:	Carlos Gonzalez-Gamella, Daniel Sanchez-Santolaya, Francisco Moraleda-Moreno and Jacobo Chaquet-Ulldemolins
Abstract:	Retrieval-augmented generation pipelines often add a reranking stage to reorder retrieved documents before prompting a large language model. While reranking has been shown to improve performance in many settings, applying it universally introduces overhead and may even degrade performance when the initial retrieval is already strong. We present the Adaptive Reranking Decision System (ARDS), a lightweight method that determines whether reranking is necessary based on retrieval signals. We have formulated this decision as a binary classification task, using local features that capture query-level uncertainty and global features that model broader patterns, including topic-cluster behavior and retriever-reranker compatibility. Among these, topic-based features emerge as strong predictors, reinforcing the value of combining micro and macro retrieval dynamics. To validate our approach, we build a large-scale benchmark across multiple datasets and open-source cross-encoder rerankers. Our findings show that universal reranking is suboptimal: ARDS reduces harmful reranks by ∼25% and increases helpful ones by a similar amount. Compared to always reranking, it improves Recall@1 by∼6% and Recall@3 by ∼10%, yielding a better effectiveness–efficiency trade-off. These findings provide a practical answer to a fundamental question in RAG pipelines: when to rerank.
Download

Paper Nr:	131
Title:	Enhancing Fake News Detection through Cross-Lingual Models: A COVID-19 Case Study
Authors:	Ezequias de Oliveira Rocha, Cláudio de Souza Baptista, André Luiz Firmino Alves and Anselmo Cardoso de Paiva
Abstract:	The proliferation of health-related misinformation on social media poses a serious threat to public well-being and the integrity of information. This paper addresses how to transfer labels effectively to a target language with limited annotations, investigating how performance degrades with lexical distance in the context of COVID-19 misinformation. Using the MM-COVID dataset comprising texts in English, Spanish, and Portuguese, we systematically evaluate five cross-lingual learning strategies: Zero-Shot Transfer, Joint Learning, Cascade Learning, a JL/CL hybrid approach, and an extended JL/CL+. The experiments employ both monolingual BERT encoders (BERT-base, BETO, and BERTimbau) and multilingual models such as XLM-R and mDeBERTa. Performance is assessed using F1 and ROC-AUC metrics, complemented by statistical significance testing. The results show that progressive adaptation strategies, including Cascade Learning and JL/CL variants, consistently outperform monolingual and zero-shot baselines, achieving F1-scores as high as 98.2 percent for Portuguese. These findings suggest that lexical distance plays a role in simple transfer setups but becomes less influential under progressive adaptation schemes.
Download

Paper Nr:	134
Title:	A Machine Learning and Process Mining Based Approach to Estimate Time to Completion and Activities in Moulds Manufacturing Processes
Authors:	Maria Rodrigues, Carlos Grilo, Lino Ferreira, Nuno Morgado, Rui Rijo, Sérgio Silva and Ricardo Martinho
Abstract:	In many industries, accurately predicting the remaining time to completion and the sequence of upcoming activities for a manufacturing process is of great importance. Such predictions help companies plan their operations to avoid bottlenecks, deviations, and non-conformities in final products or services. This challenge is especially pronounced in the mould manufacturing sector, where process executions are highly variable, reactive, and often managed in an ad-hoc manner. Because these processes depend heavily on human decisions and resource availability, traditional predictive models often fail to capture the necessary complexity. In this paper, we propose a novel approach to predict the Estimated Time to Completion (ETTC) and the sequence of remaining activities by combining Machine Learning (ML) and Process Mining (PM) techniques. Unlike standard methods, our approach does not rely on historical data alone; we distinctly integrate real-time context information to improve prediction confidence. This helps managers make more informed decisions to meet delivery deadlines. Our combined ML+PM+Context approach is shown to reduce process variability by leveraging historical process execution knowledge, leading to mould manufacturing processes that are not only predictable but also more planable and optimisable over time.
Download

Paper Nr:	135
Title:	A Scoping Review of Platform Tools for Developing LLM-Grounded Multi-Agent Systems
Authors:	Lídia Stadtlober Abrantes, Luan Kenig de Souza, Luiz Henrique Eltermann Ribeiro, Gabriel Gobbo Araújo Valera, Lucas Nardi Vieira, Guilherme Trajano, Bernardo Pandolfi Costa, Heitor Henrique da Silva, Carlos Eduardo Antônio Ferreira and Alison R. Panisson
Abstract:	This paper presents a comprehensive scoping review of available platforms and tools for the development of multi-agent systems (MAS) grounded in Large Language Models (LLMs). This review identifies and analyzes key characteristics of these platforms, such as architecture, extensibility, performance, and interoperability. This provides an overview of existing options and their suitability for various application scenarios. In addition, the primary features of each tool are detailed and compared.
Download

Paper Nr:	137
Title:	Performance Evaluation of Classification Models to Identify Architectural Elements in BIM Projects
Authors:	Eike Natan Sousa Brito and Andre Britto de Carvalho
Abstract:	Building Information Modeling (BIM) is an integrated process concept that connects project information from conception to construction, with the goal of promoting the management of all project stages. With BIM, there is an increase in the availability of structured data, enabling the application of Machine Learning techniques to improve the construction process. This work presents a performance analysis of classification models to identify architectural elements in BIM Projects. Here, data are extracted, converted into IFC (Industry Foundation Classes) format, and structured to train six classifiers: SVM, SVM with optimized hyperparameters, One-Class SVM, Decision Tree, Random Forest, and PU Learning. Three experiments were conducted to compare the performance of these models under different training and testing scenarios, allowing a detailed evaluation of their predictive behavior. The results highlight the strengths and limitations of each approach, demonstrating which classifiers achieve the highest accuracy and stability when applied to IFC-based architectural data. This study contributes to the automation of semantic enrichment in BIM workflows and supports the selection of suitable machine learning methods for classifying building elements.
Download

Paper Nr:	140
Title:	An Agentic LLM-Based System for Short-Term Crude Oil Price Forecasting
Authors:	Hind Aldabagh, Xianrong Zheng, Mohammad Najand, Sandeep Kalari and Ravi Mukkamala
Abstract:	Financial sentiment analysis is fundamental to understanding market dynamics and enabling informed trading decisions. This study contributes to the field by evaluating the performance of several state-of-the-art large language models, namely Google’s Gemini 2.5 Pro, OpenAI GPT-5, Claude Sonnet 4.5, and Grok 4, for sentiment driven prediction in the crude oil market. Using news headlines and full news articles from 2025, a few shot prompting strategy is employed to evaluate multiple prompt formulations on a carefully curated dataset of crude oil related news. The resulting prediction framework is assessed using root mean squared error (RMSE) and mean absolute error (MAE) metrics for price prediction tasks. The integration of news-based textual information yields an improvement in forecasting accuracy. Specifically, the RMSE decreases from 1.47 when relying solely on zero shot prompt data to 1.18 when incorporating news headlines and article text through few shot prompting. Similarly, the MAE improved from 1.27 to 1.03. under the same enhanced input configuration.These results demonstrate that sentiment signals extracted from news content contribute valuable supplementary information and enhance short-term crude oil price prediction.
Download

Paper Nr:	145
Title:	Light Dense-SENet: A NAS-Based Approach for Selecting a Lightweight Architecture for Thoracic Disease Classification
Authors:	Bruna Marcelly de Menezes Saraiva, Geraldo Braz Junior, Vandecia Rejane Monteiro Fernandes, Tiago Bonini Borchatt, Mackele Lourrane Jurema da Silva, Sabryna Rodrigues Araujo, Lucas Araujo Gonçalves, Rodrigo Otavio Cantanhede Costa, José Ribamar Durand Rodrigues Jr, Marcos Melo Ferreira and Joana Kuelvia de Araujo Silva
Abstract:	Diagnosing thoracic diseases using X-rays can be difficult, even for experts. In this context, deep learning can be a very helpful tool for doctors to make fast, accurate diagnoses. This work proposes classifying pulmonary infections using shallower networks. The main contribution is the use of a Neural Architecture Search (NAS) strategy with Bayesian optimization to automatically find a lightweight architecture. Experiments were done on benchmark datasets for both multiclass and binary classification of thoracic X-ray images. The results were promising: the optimized model achieved an F1-score of 95% in the multiclass task (Normal, COVID-19, and Viral Pneumonia) and 96% in the binary task (Normal vs. Pneumonia). Importantly, the selected models were substantially lighter and had fewer parameters than standard networks such as DenseNet. These findings validate that optimizing for architectural lightness is an effective strategy for deploying automated diagnostic systems in resource-constrained environments with limited hardware capabilities.
Download

Paper Nr:	146
Title:	A Comparative Study of Traditional Machine Learning and Large Language Models for Information Extraction in Public Procurement Notices
Authors:	Lucas Oliveira Belmiro, Cláudio de Souza Baptista, André Luiz Firmino Alves, Igor Silveira de Andrade, Eliane Tâmara Lima de Oliveira and Clécio Bruno Medeiros Aaraújo
Abstract:	The analysis of public procurement notices is a fundamental task for public oversight, but it remains predominantly manual due to the length and lack of structure of these documents. This work investigates the use of large-scale generative artificial intelligence models, in comparison with traditional supervised machine learning approaches, to automate information extraction from Brazilian procurement notices through sentence classification and document-level structured summarization. Experiments were conducted on an annotated dataset comprising approximately 8.5k labeled sentences extracted from 278 procurement notices, evaluating multiple generative models under a unified inference pipeline, without fine-tuning or manual data balancing, and contrasting their behavior with supervised baselines trained on the same corpus. The results show that the generative models can effectively support automated analysis, improving scalability and consistency in auditing activities, although performance limitations persist in underrepresented categories. These findings highlight the potential of generative models as a scalable and complementary alternative to traditional supervised approaches for public sector document analysis.
Download

Paper Nr:	147
Title:	Percolation-Based MRI Classification of Alzheimer's Disease Using Vision Transformers
Authors:	Sérgio Augusto Pelicano Junior, Guilherme Freire Roberto, Marcelo Zanchetta do Nascimento, Ricardo José Ferrari and Leandro Alves Neves
Abstract:	Early diagnosis of Alzheimer’s disease from structural magnetic resonance imaging remains challenging due to the subtle and heterogeneous nature of neurodegenerative changes. In this work, we propose a hybrid classification framework that integrates connectivity-based representations derived from Percolation Theory with a Vision Transformer architecture. Structural MRI volumes are preprocessed via skull stripping and intensity normalisation, and subsequently decomposed into axial, coronal, and sagittal views. Local tissue connectivity is modelled using percolation maps generated by the Gliding Box algorithm at multiple spatial scales. These maps act as binary masks applied to the original intensity images, preserving anatomical details only within structurally connected clusters to serve as input for the classification model. Classification is performed using a Data-efficient Image Transformer (DeiT-Tiny) with transfer learning. The approach is evaluated on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset using a stratified holdout validation scheme with a 70/30 training-testing split, ensuring subject-level separation. Experimental results show that fine-scale percolation (L=3) yields the most discriminative features. A multiview analysis reveals a strong dependence on anatomical orientation, with the sagittal view achieving the best performance (AUC = 0.963), slightly outperforming the baseline model trained on original images, while performance degrades in the axial view. These results demonstrate the feasibility of integrating topological connectivity modelling with Vision Transformer for Alzheimer’s disease classification and highlight the importance of anatomical orientation in percolation-based representations.
Download

Paper Nr:	154
Title:	CommentGuard: Detecting Elsagate Content through Linguistic Analysis of YouTube Video Comments
Authors:	Helena Ferreira Fernandes and Wladmir Cardoso Brandão
Abstract:	This paper presents CommentGuard, a lightweight and scalable approach for detecting harmful children-oriented content on YouTube, with a particular focus on the Elsagate phenomenon. Unlike conventional moderation pipelines that rely on computationally expensive audiovisual processing or metadata-only signals, CommentGuard leverages linguistic patterns found exclusively in user comments as an indirect but effective proxy for identifying problematic videos. We construct a curated and balanced dataset of 9,063 videos by integrating and re-annotating samples derived from two existing large-scale resources: the Elsagate Corpus and a general YouTube comment dataset used as a control group. Due to licensing and redistribution constraints associated with the original sources, the resulting derived dataset is not publicly redistributed; however, we provide a detailed and reproducible description of the construction protocol. We evaluate two representation strategies: (i) aggregated TF-IDF vectors with statistical pooling and (ii) sentence-level embeddings generated using Sentence-BERT, combined with multiple supervised classifiers (Logistic Regression, SVM, and XGBoost) under stratified cross-validation. Experimental results show that comment-only features are highly discriminative, with the best-performing models exceeding 98% accuracy. These findings indicate that comment-based analysis can provide an interpretable and computationally efficient layer for large-scale moderation aimed at protecting children on digital platforms.
Download

Paper Nr:	176
Title:	Efficient Money Laundering Detection through Graph-Based Prioritization
Authors:	André da Costa Silva and Paulo André Lima de Castro
Abstract:	Anti-Money laundering (AML) transaction monitoring must prioritize a small fraction of high-volume transfers for investigation under strict capacity constraints. Standard threshold metrics (e.g., AUC, F1) do not reflect this setting. We propose a two-stage pipeline that (i) learns a transaction-level risk ranking over a directed transactional multigraph using a graph neural network (GNN) encoder with an edge decoder and (ii) transforms the inspected top-k% prefix into investigable cases by extracting weakly connected components and adding controlled contextual evidence via subgraph induction. We evaluate four GNN encoders on outof-time AMLSim splits at two scales (AML100k and AML1M) with budget-aware metrics (Recall@k% and inspections-per-illicit-transaction) and case-level metrics (case coverage, induced-context purity, and intracase completeness, CR@k) to quantify the trade-off between evidence recovery and contextual noise. Among the evaluated GNN encoders, GraphSAGE achieves Recall@1% = 0.877 on AML100k and 0.870 on AML1M, requiring 7.94 inspections per illicit transaction at k=1%, while sustaining approximately 0.74 million edges/s on AML1M. Results show rapid saturation beyond the first percentiles and an explicit coverage–purity tradeoff as k increases, supporting budget-aware tuning for operational AML decision support.
Download

Paper Nr:	225
Title:	A Lifecycle-Oriented Data Governance Framework for Large Language Models
Authors:	Stefano Luppi Spósito, Muhtasim Azad, Leonardo Machado Santos, Fábio Lúcio Lopes de Mendonça and Edna Dias Canedo
Abstract:	Context: Large Language Models (LLMs) introduce intertwined risks related to privacy, security, bias, and regulatory compliance due to their data-intensive and lifecycle spanning characteristics. Goal: This study aims to design a lifecycle oriented data governance framework that operationalizes ethical principles and legal obligations (LGPD/GDPR) through auditable organizational and technical controls tailored to LLM environments. Method: Following a Design Science Research approach, we conducted a structured literature synthesis, derived governance requirements, designed a directive based framework, and evaluated it using normative alignment analysis, threat-to-control mapping, and scenario based walkthroughs. Results: The resulting framework integrates five transversal directives: bias mitigation, ethics, security and privacy, data quality, and regulatory compliance applied across all LLM lifecycle phases. These directives are operationalized through defined organizational roles, structured governance processes, and evidentiary mechanisms such as dataset versioning, lineage tracking, logging, DPIA/PIA documentation, and unlearning pathways. Conclusions: The framework bridges technical safeguards, organizational accountability, and regulatory alignment, enhancing auditability and supporting governance readiness for LLM adoption in regulated and high-impact organizational contexts.
Download

Paper Nr:	246
Title:	A Soft Computing Framework for Iris Liveness Detection Using Multi-Filtering and Texture Descriptors
Authors:	José Fabio Saraiva de Oliveira, Monique Simplicio Viana, Rodrigo Colnago Contreras, Rodrigo Capobianco Guido, Önsen Toygar and Mubeen Tajudeen
Abstract:	Iris biometric recognition remains vulnerable to presentation attacks such as printed images and textured contact lenses. This work proposes a soft computing framework combining multi-filtering and texture descriptors for liveness detection. The methodology applies data augmentation, Region of Interest detection (ROI) extraction, and the sequential application of Gaussian, High-Pass, and Guided filters, followed by the image enhancement algorithm Contrast Limited Adaptive Histogram Equalization (CLAHE). Feature extraction utilizes descriptors such as Co-occurrence of Adjacent Local Binary Patterns (coALBP), Local Phase Quantization (LPQ), and statistical Dense Scale-Invariant Feature Transform (sDenseSIFT), mapped through statistical functions. Classification is performed using Singular Value Decomposition (SVD) for dimensionality reduction, followed by Support Vector Machines (SVM). Evaluated on the LivDet-Iris 2013, 2015, and 2017 benchmarks, the framework achieves competitive results, attaining the lowest Average Classification Error (ACE) on the 2013 and 2017 editions, while maintaining solid performance in 2015.
Download

Paper Nr:	259
Title:	A Hierarchical Framework for Enterprise-Scale Aspect Extraction from Code-Switched E-Commerce Reviews: A Comparative Deployment Study
Authors:	Valiant Lance D. Dionela, Fatima Kriselle S. Dy, Robin James M. Hombrebueno, Aaron Rae M. Nicolas, Charibeth K. Cheng and Raphael W. Gonda
Abstract:	E-commerce platforms in multilingual markets struggle to extract actionable insights from code-switched reviews. Filipino platforms such as Shopee process millions of “Taglish” (Tagalog-English) reviews annually without automated tools for aspect-level analysis. This study presents a decision support framework for aspect extraction in low-resource, code-switched environments. A Hierarchical Aspect Framework (HAF), developed through topic modeling triangulation, organizes feedback into 4 general and 21 specific business categories. To support technology selection for enterprise deployment, four distinct approaches are compared: (1) Rule-Based systems, (2) Generative Large Language Models (Gemini 2.0 Flash), and (3) two fine-tuned Gemma-3 models trained using different annotation strategies. Evaluated on 10,510 real-world reviews, the generative LLM achieves an F1 score of 91%, excelling at implicit aspects, while rule-based systems offer up to 70% cost reduction for high-volume tasks. Fine-tuned models show limited viability due to capacity constraints. A cost-benefit analysis further indicates that hybrid deployment strategies can reduce manual review analysis costs by approximately 85% while maintaining enterprise-grade accuracy. The proposed framework enables a scalable, linguistically adaptive customer intelligence solution, with applications in product development, customer service, and competitive analysis across Southeast Asian e-commerce.
Download

Paper Nr:	262
Title:	AI in Asset and Liability Management: ALM - Trends, Challenges and Opportunities in Financial Risk Management
Authors:	Gisele Carolina Almeida, Edson Ribeiro dos Santos Junior and Paulo André Lima de Castro
Abstract:	This study presents a systematic literature review on the application of Artificial Intelligence (AI) techniques in Asset and Liability Management (ALM) within financial institutions. The objective is to map predominant approaches and identify research opportunities in the field of AI in finance, with the aim of improving the management of financial risks, including market, credit, liquidity, operational, and legal risks. The findings indicate a rapid evolution in this field, particularly through the use of Deep Learning (DL), Reinforcement Learning (RL), ensemble algorithms, and Natural Language Processing (NLP), which consistently demonstrate superiority over traditional methods in complex forecasting and optimization tasks. A notable trend is the increasing emphasis on interpretability through Explainable Artificial Intelligence (XAI), which seeks to render algorithmic decisions understandable to humans, as well as the integration of multiple data sources. Nevertheless, significant challenges remain, including data quality issues, the trade-off between accuracy and explainability, the need for robust validation, and ethical considerations. Overall, while AI offers transformative potential for ALM, its effective and responsible implementation depends on continued advancements in areas such as scenario generation, XAI techniques, and adaptive models, along with close collaboration among academia, industry, and regulatory bodies.
Download

Paper Nr:	283
Title:	A Workflow for Structured Information Extraction from Public-Sector PDF Documents Using a Small Language Model
Authors:	Thaís B. M. B. de Sousa, Paulo V. A. Fabrício, Paulo A. T. Lima, João G. I. Braga, Rafael R. Pereira, Tales P. Nogueira, Wendley S. da Silva, Antonia M. A. da Silva, Rossana M. C. Andrade and Alexandre S. Cialdini
Abstract:	Public administration processes frequently require the validation of large volumes of heterogeneous PDF documents, making manual analysis time-consuming, error-prone, and difficult to scale. This paper addresses this challenge in the context of civil servant promotion requests, where specific information (e.g. personal identification data, legal references, dates, and monetary values) must be accurately extracted and validated across multiple documents. We propose a structured workflow for the extraction, structuring, and validation of information from governmental PDF documents under computational constraints. The approach utilizes a Small Language Model for information extraction and semantic structuring, through a few-shot prompt strategy. Extracted data are organized into JSON files and evaluated against predefined business rules to automatically generate a technical recommendation for the request’s approval or denial. The methodology is evaluated through a real-world case study using official public administration documents, yielding an average processing time of 2 min 24 s and mean accuracy of 99.9%. The results indicate that Small Language Models provide a viable trade-off between performance and resource consumption, supporting document automation in constrained environments.
Download

Paper Nr:	293
Title:	Designing a Structured AI Change Process for Enterprise Systems
Authors:	Simon Schou and Torben Tambo
Abstract:	Artificial intelligence (AI) is increasingly positioned as a core enabler of digital transformation within enterprise information systems (EIS). However, many organizations struggle to scale AI beyond isolated pilots into stable, enterprise-wide integration. This study examines how organizations can design a structured AI change process that supports successful integration of AI into internal IT systems and operational workflows. Drawing on a qualitative single-case study in a Danish manufacturing company, the research identifies five interconnected barriers to AI adoption: lack of strategy, change management, training, communication, and resources. Rather than treating AI adoption as an individual acceptance issue, the findings reveal systemic enablement gaps at the enterprise level. Based on these insights, the paper proposes the Enterprise AI Enablement Loop (E-AIEL), an iterative model integrating strategic alignment, governance embedding, capability development, workflow integration, and feedback-based learning. The study contributes to the EIS field by bridging technology acceptance theory, AI implementation research, and structured change governance, offering a design-oriented framework for moving from AI experimentation to institutionalized, enterprise-wide value creation.
Download

Paper Nr:	315
Title:	An Adaptive Multi-Agent System with Data-Driven Heuristic Orchestration for Solving 1D/2D Bin-Packing Problems
Authors:	Sam Guessoum, Chun-Kit Ngan and Rolf Bardeli
Abstract:	This paper introduces an advanced Multi-Agent System (MAS) featuring a five-agent architecture designed to tackle bin packing problems (BPPs). This framework seamlessly integrates machine learning-driven heuristic selection, adaptive search termination, and unified dimensionality handling. Our end-to-end MAS architecture closes the performance gap between heuristics, the tuning overhead of metaheuristics, and the computational burden of exact methods by coordinating five specialized agents. Each agent is classified using established AI agent taxonomy. We rigorously evaluate our framework against five heuristics, three metaheuristics, and one exact method (Branch and Bound (B&B)) across over 3,000 instances from 12 diverse datasets. The MAS achieves a 9% average improvement in bin utilization while exhibiting lower variability across instances over all heuristics and metaheuristics. It also proves 10× faster than B&B, consistently maintaining an optimality gap of ≤2% while exhibiting low variability across instances and reducing bin usage. Our statistical validation confirms significant outperformance against heuristic (α = 0.05, p-values 0.0001-0.0011) and metaheuristic baselines (p-values 0.0038-0.0061) in optimality gap and bin usage. Crucially, there is no significant difference from B&B (p=0.0674) despite being 10× faster. We also develop a web-based Proof-of-Concept prototype, providing an interactive environment for solving both 1D and 2D BPPs.
Download

Paper Nr:	329
Title:	The Storyteller in the Model: Narrative Pattern Inheritance, Escalation Dynamics, and Alignment Governance in LLMs
Authors:	Adam Rigby, Raz Saremi, Azadeh Sohrabinejad and Mehdi Rahimi
Abstract:	LLMs are trained predominantly on human-authored text, yet the structural and narrative conventions embedded in that text are rarely examined as a source of systematic behavioral influence, or as a governance risk in deployed systems. This paper considers whether the storytelling patterns inherent in published human writing, including archetypal roles such as protagonist, antagonist, and underdog, as well as tension-and-resolution narrative arcs, are absorbed during training and subsequently surface in LLM outputs, causing responses to drift toward unexpected, adversarial, or rhetorically enticing behaviors over extended interactions. Through a systematic literature review and cross-paper analysis of recent empirical studies on LLM alignment, persona dynamics, emergent misalignment, and user interaction patterns, we observe evidence bearing on this hypothesis. The findings reveal three key patterns. First, LLMs reproduce statistical patterns from their training data rather than reasoning independently. Second, measurable latent traits, including sycophancy and deceptiveness, emerge reliably across unrelated prompts. Third, fine-tuning on a narrow narrative task can produce unintended behavioral changes well beyond that task. Furthermore, evidence suggests that persuasive, narrative-style outputs are among the most common LLM products in real-world usage, amplifying these risks. Narrative drift constitutes an unmonitored escalation pathway in deployed AI systems, one that evades discrete-incident detection mechanisms and requires dedicated monitoring instruments. Collectively, the evidence reviewed is consistent with the hypothesis that hidden narrative patterns in training data exert a statistically traceable and practically significant influence on LLM behavior, and this influence intensifies across longer conversational exchanges, posing underappreciated risks for reliability, alignment, and governance of deployed systems. We argue that mitigating these risks requires concrete control instruments: narrative-archetype classifiers, structural drift metrics, and longitudinal escalation frameworks capable of detecting and intervening before behavioral divergence undermines system reliability.
Download

Paper Nr:	330
Title:	Integrating AI Agents into Business Process Automation: Evidence from Customer Success Operations
Authors:	Monique R. Moraes, Gabriel Galvão, Martin Augusto Gagliotti Vigil, Olga Yevseyeva, Analúcia Schiaffino Morales, Heinz Felipe Cavalcante Rahmig, Iwens Gervásio Sene-Junior and Alison R. Panisson
Abstract:	Customer Success teams play an important role in ensuring that organizations extract long-term value from digital products and services. However, many customer success activities involve repetitive documentation tasks, fragmented information management, and manual coordination across multiple systems, limiting the time available for strategic client engagement. This paper presents an agentic architecture that integrates Large Language Model (LLM) agents with a workflow orchestration platform to support knowledge-intensive customer success processes. The proposed system automates meeting documentation, task generation, and account handover preparation while maintaining human oversight through human-in-the-loop mechanisms. The architecture was implemented in a real organizational environment and evaluated through two agent-assisted workflows. Results indicate significant reductions in operational workload, with substantial time savings in meeting documentation and task management activities. Beyond the specific case study, the findings illustrate how agent-based automation can extend traditional workflow automation by enabling adaptive and context-aware business processes while preserving governance and human supervision.
Download

Short Papers

Paper Nr:	22
Title:	Integration of Fuzzy Logic and Cognitive Computing in Document Management and Payment Systems
Authors:	Elchin Aliyev, Ramin Rzayev, Elchin Mahmudov and Ayten Rahmanova
Abstract:	The growing volume of digital documents and financial transactions requires intelligent decision-support mechanisms capable of operating under uncertainty. This paper proposes an approach integrating fuzzy logic and cognitive computing for document classification and risk-oriented analysis in enterprise document management and payment systems. The method combines expert knowledge aggregation, fuzzy inference modelling, and neuro-fuzzy learning to support automated confidentiality assessment and document routing. The framework was tested using enterprise platforms developed by SINAM Ltd., including the SESDA document management system and the SEPS electronic payment system. Experimental results show that the approach improves document classification accuracy by approximately 15–20% compared with conventional rule-based methods while reducing the influence of subjective expert judgement.
Download

Paper Nr:	57
Title:	Towards Reliable AI in Public Administration: An Analysis of LLM-Generated Legal Opinions
Authors:	Pedro Bilar Montero, Felipe Amadori Machado, Fernando Soso Girardi, Gabriel Rodrigues Silva, Thiago Paulo Both, Gisane Lanes de Almeida, Vinícius Maran and Alencar Machado
Abstract:	The analysis of administrative procedures in public agencies, particularly in specialized domains like animal health, presents significant challenges due to the volume of heterogeneous data, complex regulations, and the need for consistent, high-stakes decision-making. The emergence of Large Language Models (LLMs) offers a promising avenue for supporting and automating these resource-intensive tasks, a topic of growing interest in both public administration and applied AI research. This paper proposes a novel architecture for an intelligent assistant designed to aid public sector analysts by integrating a core LLM with a Retrieval-Augmented Generation (RAG) framework, ensuring that generated analyses are grounded in specific legal and normative documents. To validate this approach, a prototype of the architecture’s core generative engine was evaluated in a real-world scenario: the analysis of administrative cases within the Administrative Review Board (JAP) of the Secretariat of Agriculture, Livestock, and Rural Development of Rio Grande do Sul (SEAPI-RS). In this foundational experiment, the system generated technical opinions for fourteen cases, which were then systematically compared against the final decisions of human analysts. The results validate the viability of using an LLM as the engine for the proposed assistant, demonstrating stable decision-making accuracy. However, the analysis also revealed a failure mode where the model could be misled by sophisticated but incorrect legal arguments, underscoring the necessity of the RAG framework for grounding and reliability. Furthermore, the findings highlight the importance of hyperparameter tuning to align the model’s output style with the concise and deterministic nature required in high-stakes administrative domains. This study provides the empirical basis to proceed with the full implementation of the proposed architecture.
Download

Paper Nr:	59
Title:	Towards a Software Architecture to Integrate Case Based-Reasoning and Large Language Models for Fishery Product Certification
Authors:	Gabriel Vieira Casanova, Pedro Bilar Montero, Gabriel Rodrigues da Silva, Thiago Paulo Both, Matheus Friedhein Flores, Alencar Machado and Vinicius Maran
Abstract:	The administrative processes within regulatory platforms, such as Brazil’s National Fishery Industry Platform (PNIP), are often complex and laden with repetitive tasks, leading to inefficiencies and potential errors. The Legal Origin Accreditation Certificate (CAOL) is a critical process within PNIP, involving multiple steps, Normative Instructions (NIs), and dispatches. This paper proposes the development of an intelligent agent to assist in the CAOL submission. The agent leverages the Model Context Protocol (MCP) to create a standard interface between a local Large Language Model (LLM) and the PNIP platform. It integrates Retrieval-Augmented Generation (RAG) to provide contextual assistance by retrieving relevant regulatory instructions and procedural explanations. Furthermore, Case-Based Reasoning (CBR) is employed to suggest progress percentage for new CAOL forms and dispatches based on historical similar cases (see Sec. 3.4.1 for the computation used). This paper outlines the architecture of this integrated solution, discusses potential evaluation methodologies, and explores its implications for reducing bureaucratic overhead and improving compliance within Brazil’s fishery sector.
Download

Paper Nr:	60
Title:	A Weakly Supervised Machine Learning Framework for Classifying Fishing Trips from VMS Data
Authors:	Leonardo Petta do Nascimento, Murillo D'Almeida Couto de Azevedo, Alex Souza Lira, Alencar Machado and Vinícius Maran
Abstract:	Sustainable fisheries management relies on accurate monitoring of vessel activities. While Vessel Monitoring Systems (VMS) provide large-scale positioning data, segmenting trajectories into fishing behaviors remains a challenge due to the scarcity of labeled datasets for training models. To bridge this gap, this paper proposes a Weakly Supervised Machine Learning framework applied to purse seine fisheries. We developed an Expert System based on domain heuristics, integrating speed, dwell time, and spatial history, to generate synthetic labels for 65,098 VMS records from the Brazilian fleet (2021–2025). Subsequently, we trained supervised models (Random Forest, LightGBM, XGBoost, and Gradient Boosting) to learn the latent behavioral patterns using only kinematic and spatiotemporal features, ensuring strict prevention of data leakage. LightGBM achieved the best performance with 92% accuracy, followed closely by Random Forest with 91% on a strict time-based test set (2025 season). Feature importance analysis revealed that kinematic variables are the primary predictors, confirming that the expert rules are grounded in a distinct physical signature of the fishing operation. These results validate the framework as a robust and scalable approach for fishery monitoring, capable of effectively distinguishing fishing effort without reliance on intensive manual annotation.
Download

Paper Nr:	62
Title:	Safeguarding Big Data Question-Answering Systems with a Two-Layer Input Guardrails
Authors:	Leonardo Mauro Pereira Moraes, Enzo Baraldi Onofre and Cristina Dutra Aguiar
Abstract:	Big data question-answering systems process natural-language questions over large text datasets. They rely on advanced question-answering techniques to understand and interpret user queries, capture intent, and generate relevant answers, while handling high query throughput through document-based knowledge bases. BigQA is a big data question-answering architecture that meets these requirements but may be susceptible to critical vulnerabilities. To address the challenge of reliably finding a compliant answer, we propose a security-first architecture, Safe-BigQA, that extends BigQA by embedding guardrails within the querying layer. Our core contribution is a two-layer guardrails algorithm for input validation that combines a fast, small language model as a high-precision filter with a more accurate, but slower, large language model as a fallback for complex cases. We validated our proposal through extensive jailbreak experiments across three datasets, comparing our algorithm against six state-of-the-art models. The results showed that the proposed algorithm achieved the highest performance in jailbreak detection (F1 score up to 62.87%) on key safety benchmarks, representing an increase of 9.64% over the standalone LlamaGuard. This result demonstrates that the architectural design of Safe-BigQA and its two-layer guardrails algorithm significantly enhances the reliability of big data question-answering systems in high-stakes environments.
Download

Paper Nr:	64
Title:	Clustering-Driven AI Optimization of Radial Cable Infrastructure in Offshore Wind Farms
Authors:	Italo Firmino da Silva, Tamiris Grossl Bade, Wilian Cominn, Telles Brunelli Lazzarin, Lenon Schmitz and Alison R. Panisson
Abstract:	Offshore wind farms play a key role in the global transition to renewable energy, benefiting from stronger and more consistent wind conditions than onshore sites. However, the high costs associated with installation and maintenance require efficient design strategies that jointly address energy yield and infrastructure expenses. While prior research has primarily focused on optimizing turbine placement to mitigate wake effects and maximize energy production, this work addresses the complementary challenge of optimizing electrical cable infrastructure. We propose a clustering-based algorithm for designing cost-effective radial cable layouts, integrated into an existing AI-driven wind farm design framework. The method groups turbines and computes optimized cable routes to minimize total cable length and transmission losses while satisfying radial topology constraints. Experimental results demonstrate that the proposed approach effectively supports the design of economically and technically efficient wind farm cable networks, enhancing integrated design frameworks with a practical tool to improve the overall viability of offshore wind projects.
Download

Paper Nr:	69
Title:	From Documents to Decisions: An Enterprise RAG Architecture for Technical Tender Preparation
Authors:	Marco Claps, Giovanni Simonini and Giorgio Zucchi
Abstract:	This work, arising from a real case study, addresses a central challenge in facility services sector: project designers must navigate large volumes of regulatory and technical documents to prepare and to write tender submissions, a process that is slow, error-prone, and costly. To overcome these limitations, we present a modular retrieval-augmented generation (RAG) system specifically engineered and created for regulated industrial workflows. The proposed system are able to combine hybrid retrieval, cross-encoder re-ranking, and GPT-based reasoning within a secure, closed environment. Evaluated on more than 500 heterogeneous real documents, written in Italian language, and tested with expert users from the company, the system significantly improves retrieval accuracy, reduces response time by over 50% compared to TF-IDF, and increases organizational project throughput by 12%. Our findings show that domain-grounded retrieval, coupled with transparent references and workflow integration, provides a practical and scalable path to reliable AI-assisted tender design.
Download

Paper Nr:	73
Title:	Mutual Information Maximization for Data Enrichment and Enhancement of Predictive Models in Software Effort Estimation
Authors:	Leonardo de Oliveira Campos, Alisson Marques da Silva and Mirela Teixeira Cazzolato
Abstract:	Given the growing complexity of software projects, how can we improve effort estimation to increase the accuracy of predictive models and assist in development planning? This paper proposes InfoMIMIC, an approach that combines data preprocessing techniques with the Mutual Information Maximization for Input Combination (MIMIC) algorithm, focusing on generating derived variables that enrich the databases used for estimation. The experimental evaluation reveals improvements in metrics such as RMSE and MAE, indicating that the approach reduces errors in the predictions of Linear Regression, Ridge, and Random Forest algorithms. InfoMIMIC represents a promising alternative for advancing effort estimation in software engineering, opening new possibilities for investigation in database enrichment.
Download

Paper Nr:	114
Title:	One Look to Rule Them All: Leveraging YOLOv11 for Robust Traffic Sign Detection
Authors:	Maria Eduarda Ribeiro da Rocha and Mateus Coelho Silva
Abstract:	The rapid advancement of computer vision techniques has revolutionized autonomous driving, making real-time traffic sign detection a critical component for safe vehicle navigation. Efficiently identifying these regulatory cues is essential for ensuring compliance with road rules and minimizing accident risks in complex driving environments. This article investigates the application of the YOLOv11n model for traffic sign recognition in the context of autonomous navigation, analyzing the relationship between detection accuracy and computational efficiency. The methodology used a customized dataset of Brazilian signs, trained for 200 epochs with automatic optimizer adjustment and a learning rate of 0.0005. The experimental results demonstrate high perceptual performance, reaching 0.912 mAP@50 and 0.803 mAP@50–95, with stable convergence and no signs of overfitting. Computational performance analysis recorded an average latency of 16.3 ms, validating the model’s viability for applications that require low latency and fast responses. The study highlights YOLOv11n’s potential to integrate assisted navigation systems, emphasizing its balance between visual robustness and processing agility.
Download

Paper Nr:	115
Title:	An Industrial AI Agent for Automated Parameter Optimization: A Reinforcement Learning Approach to Engine Calibration
Authors:	Rodrigo Luan Ferrazza, Afonso Sales and Karina Ruschel
Abstract:	The manual calibration of control parameters in complex industrial machinery, particularly in the automotive sector, is a highly heuristic, time-consuming, and expert-dependent process. This bottleneck challenges the scalability of modern manufacturing and maintenance workflows. This paper proposes an Intelligent Decision Support System (IDSS) driven by Reinforcement Learning (RL) to automate the generation of electronic fuel injection maps, effectively transforming a complex physical calibration task into an autonomous optimization problem. The proposed architecture couples a Q-Learning agent-acting as the cyber-intelligent core-with a simulated digital twin of the engine dynamics, derived from over 380 real-world operational datasets. Unlike traditional model-based control, the agent learns an optimal control policy directly from sensor data (RPM, Manifold Pressure, Lambda) through interaction with the simulated environment. The system encodes engineering objectives, such as stoichiometric efficiency and operational stability, into a reward function that guides the agent toward convergence. Experimental results demonstrate that the system achieves stable policy convergence after approximately 30,000 episodes, producing robust and physically coherent control maps. These findings validate the feasibility of employing Industrial AI agents to automate critical parameter tuning processes, establishing a foundation for adaptive, data-driven calibration in embedded enterprise systems.
Download

Paper Nr:	128
Title:	Automated Video Analysis for Process Improvement in Lean Management
Authors:	Arbnor Bekiri, Maja Spahic-Bogdanovic, Hans Friedrich Witschel and Barbara Re
Abstract:	Traditional Lean Management relies primarily on manual, experience-driven observation to detect waste (muda), making waste analysis time-consuming, subjective, and challenging to scale. Although recent advances in computer vision enable automated process monitoring, existing approaches rarely translate visual observations into interpretable Lean waste categories. An automated, video-based system is introduced for identifying Lean waste in manual assembly processes. This method integrates computer vision and machine learning to extract object- and human-focused features from video data and then classifies these features into Lean waste categories using explicit logic. To improve comprehension, large language models produce structured natural-language explanations that link detected instances of waste to Lean principles. The framework is assessed with a publicly available video dataset documenting a wooden box assembly process, which includes manually annotated ground truth. The results indicate high precision for waste types with clear visual or temporal characteristics, such as waiting, and moderate performance for more ambiguous categories, including motion. These findings indicate that automated video analysis can reduce reliance on subjective manual observations and offer understandable decision-support results that assist human experts in Lean process improvements, without replacing their judgment.
Download

Paper Nr:	151
Title:	Model Comparison for Adversarial Attack Detection in EEG-Based Brain-Computer Interfaces
Authors:	Beatriz Conceição da Costa, Giancarlo Lucca, Lizandro de Souza Oliveria, André Riker and Bruno L. Dalmazo
Abstract:	Brain-Computer Interfaces (BCIs) are becoming increasingly common in assistive and consumer technologies, allowing users to control external devices through brain signals. However, the reliability of these systems can be compromised by adversarial attacks that subtly alter input data to mislead machine learning models. This paper evaluates how vulnerable EEG-based BCI classifiers are to such attacks. Using the Foolbox framework, we simulated four common adversarial methods (FGSM, PGD, DeepFool, and Carlini & Wagner) and measured their effects on EEGNet performance. The experiments showed that, after the attacks were applied, the classification performance was reduced to about half of its original level, with attack success rates ranging from 75.2% to 100%. To mitigate these threats, we tested three traditional detection models (Random Forest, SVM, and KNN), achieving up to 83% accuracy in identifying adversarial attacks. The results confirm that EEG-based BCIs are highly susceptible to small signal perturbations and highlight the need for more robust defense and detection strategies to ensure their security and reliability.In addition, an Adaptive Neuro-Fuzzy Inference System (ANFIS) was evaluated as an uncertainty-aware detection mechanism.
Download

Paper Nr:	166
Title:	RegNet to Ophthalmic Segmentation Architecture: ROSA
Authors:	Thiago Paiva Freire, Geraldo Braz Júnior, João Dallyson Sousa de Almeida and Tiago Bonini Borchartt
Abstract:	This work proposes ROSA (RegNet to Ophthalmic Segmentation Architecture), a deep neural network architecture for automatic segmentation of the optic disc (OD) and optic cup (OC) in fundus images, aiming to aid early glaucoma diagnosis. The methodology is based on neural architecture search (NAS), using the principles of the RegNet search space and adapting them to a U-Net architecture with an encoder body, a decoder body, and skip connections. The optimization was conducted through a three-stage Grid Search to determine: (i) block configurations (residual, convolutional, or attention-based), (ii) the number of block repetitions, and (iii) the integration of the SimAM attention mechanism within skip connections. Trained on a combination of the ORIGA and REFUGE2 datasets using Combo Loss, ROSA achieved promising results with an average Dice coefficient of 95.61% for the OD and 86.95% for the OC. The model’s robustness and generalization were further validated on the DRISHTI-GS dataset, achieving Dice coefficients of 94.62% for the OD and 85.11% for the OC, while maintaining high performance across other clinical samples.
Download

Paper Nr:	174
Title:	Industrial Accident Prevention Using Computer Vision and Edge Computing
Authors:	Frederico Augusto Cardoso Diniz and Ricardo Augusto Rabelo Oliveira
Abstract:	This research explores a computer vision solution in a proof-of-concept format, evaluating a computer vision system for industrial safety on edge hardware (Raspberry Pi 4 and 5) under the guidelines of NR-12. Instance segmentation using YOLOv11 was compared with the MediaPipe landmark mechanism to identify upper limb intrusion into risk zones. Based on ISO 13855, the results show that latency directly affects human movement displacement, revealing that YOLOv11 prioritizes contour fidelity, while MediaPipe enables critical reaction times for real-time accident prevention in Industry 4.0.
Download

Paper Nr:	187
Title:	Multiclass Classification of Endometriosis in Laparoscopic Images Using Data Augmentation Techniques and Deep Learning
Authors:	Igor R. B. Estrela, Juliany P. Costa, Anselmo C. Paiva and Aristófanes C. Silva
Abstract:	Endometriosis is a chronic gynecological disease whose definitive diagnosis generally relies on laparoscopy. Although deep learning approaches have advanced computer-aided diagnosis, most existing studies remain limited to binary classification, while multiclass differentiation of laparoscopic lesions is still underexplored. This limitation arises from the scarcity of annotated data, severe class imbalance, and high visual variability of endometriotic lesions. In this work, we investigate the feasibility of multiclass classification of endometriosis in laparoscopic images using the public GLENDA dataset. We propose a unified experimental pipeline that evaluates multiple deep learning architectures under consistent training conditions, employing five-fold stratified cross-validation and systematic data augmentation strategies. In addition, a progressive training scheme is adopted, in which models are initially adapted to the laparoscopic domain through a binary classification task and subsequently refined for multiclass prediction. The results indicate consistent gains in stability and generalization. In the proposed three-class configuration, the ConvNeXt architecture achieved the best performance, reaching an average F1-score of approximately 80% and an accuracy of 85%, demonstrating the potential of GLENDA for more informative multiclass applications.
Download

Paper Nr:	191
Title:	Urban Tree Species Identification Using Deep Metric Learning-Based Image Retrieval
Authors:	Anderson Silva, Marcos Farias, Joana Silva, Levi Santos, Patrick Araujo, Maxwell Silva, Mario Freitas, João de Souza, Filipe Belfort, Davi Viana, Italo Santos, João de Almeida, Anselmo de Paiva, Aristófanes Silva, Laura Castro and Luan Sousa
Abstract:	The analysis of urban tree species from images acquired in outdoor environments is a challenging task due to high visual variability, background clutter and limited labeled data. In this context, Content-based Image Retrieval (CBIR) can be a promising alternative to conventional classification approaches by enabling similarity-based analysis rather than categorical prediction. This work proposes a CBIR approach based on Deep Metric Learning (DML), in which Convolutional Neural Networks (CNNs) are employed as deep feature extractors to embed full-tree images into a discriminative latent space. The learned embeddings are optimized using N-pair Loss, allowing similar tree species to be mapped closer together while separating dissimilar ones. Among the evaluated models, EfficientNetB0 achieved the best overall performance, reaching a Top-1 Accuracy of 94.93 ± 6.87\%, Prec@10 of 95.57 ± 5.85\% and mAP@10 of 94.95 ± 6.81\%, highlighting the retrieval capacity of CNNs feature extractors for DML in this scenario. This study contributes by demonstrating the applicability of DML for tree species analysis even in noisy real-world conditions and by providing a comprehensive comparison of deep feature extractors within this CBIR approach.}
Download

Paper Nr:	194
Title:	A Short Survey on Hybrid and Traditional Evaluation Methods for LLM Outputs
Authors:	Gabriel Rudan Sales Matos, Jose Wellington Franco da Silva, Amanda Drielly Pires Venceslau and Jose Antonio Fernandes de Macedo
Abstract:	The rapid adoption of Large Language Models (LLMs) across diverse domains has intensified the need for reliable and interpretable evaluation methodologies. Despite widespread use of automatic metrics, assessing the quality, correctness, and usefulness of LLM-generated responses remains a challenging, unresolved problem. This paper presents a focused analytical review of evaluation methods for LLM outputs, synthesizing evidence from 44 primary studies published between 2023 and 2026. The review systematically analyzes traditional automatic metrics, embedding-based approaches, LLM-based evaluators, and human evaluation protocols, highlighting their respective strengths and limitations. The results show that classical metrics such as BLEU and ROUGE remain prevalent, while LLM-as-a-Judge approaches have gained significant traction for semantic and open-ended tasks. The analysis further reveals a clear trend toward hybrid evaluation frameworks that combine automatic, human, and LLM-based methods to balance scalability and analytical depth. This work identifies emerging patterns, open challenges, and research opportunities, contributing to a clearer understanding of the state of the art in LLM evaluation.
Download

Paper Nr:	196
Title:	When AI Meets Software Maintenance: Insights from Brazilian Developers
Authors:	Gabriel da Silva Freitas and Bruno Gadelha
Abstract:	Software maintenance is widely recognized as the most effort- and cost-intensive phase of the software development lifecycle. In recent years, Artificial Intelligence (AI) tools, particularly Large Language Models (LLMs) and AI coding assistants, have increasingly been adopted to support maintenance activities. However, empirical evidence on how developers integrate these tools into daily maintenance work remains limited, especially in industrial contexts from developing regions. This paper investigates how AI-based tools and traditional resources are used in software maintenance within Brazilian software development companies. We conducted a personal opinion survey with 47 industry professionals, collecting both quantitative and qualitative data. Results indicate that AI tools are mainly perceived as mechanisms to accelerate maintenance tasks, automate repetitive activities, and support code understanding. Traditional resources still remain essential for deeper technical understanding and confidence in problem resolution. Overall, the results highlight a complementary relationship between AI-based tools and traditional maintenance resources.
Download

Paper Nr:	197
Title:	Automatic Quality Assessment of Anterior Segment Ophthalmic Images Using Deep Learning Ensemble
Authors:	Emily Gabriela Costa Ribeiro, Jhones de Sousa Soares, João Marcello Mendes Moreira, João Dallyson Sousa de Almeida, Elaine de Paula Fiod Costa and Aristófanes Corrêa Silva
Abstract:	Automatic quality assessment is a key step in ophthalmic computer-aided diagnosis, especially in telemedicine and large-scale screening, where acquisition conditions are heterogeneous. Portable devices enable decentralized imaging in remote or underserved regions, but variations in focus, illumination, and operator expertise often degrade image quality, making quality control necessary before diagnosis. This study investigates automatic quality assessment on a private dataset of 810 anterior segment images captured with a portable device and annotated by specialists into three classes (Good, Usable, Rejected). A transfer-learning pipeline was evaluated with patient-grouped stratified 5-fold cross-validation and weighted loss to address class imbalance. Several convolutional and attention-based architectures were compared, and performance was improved with unweighted probability-averaging ensembles assessed through a systematic ablation study. On the independent test set, the best ensemble achieved 0.788 accuracy, 0.791 macro F1-score, and 0.683 Cohen’s κ. Normalized confusion matrices and Grad-CAM visualizations were used to analyze errors and interpret model behavior.
Download

Paper Nr:	218
Title:	Cross-Border e-Commerce Customs Risk Management: Exploring the Potential of Linking Digital Product Passport Data, X-Ray Scanned Images, and AI
Authors:	Boriana Rukanova, Justin Dauwels, Ger C. M. Koomen, Yao-Hua Tan, Susana Wong Chan, Frank Janssens and Toni Männistö
Abstract:	Cross-border e-commerce is continuously growing with rapid speed which poses issues for authorities to monitor and control the large volumes of goods entering the EU via postal and express services. Digital Product Passports (DPP) are seen as a digital tool that can enable the e-commerce monitoring, however what roles DPPs can play is not yet fully understood. In this research, based on real-life piloting with scanned x-ray images of 10 packages with textiles and toys and based on product data, we gained insights and defined further research directions for exploring further the potential of DPPs and AI and scanned images for customs risk management in the context of cross-border e-commerce.
Download

Paper Nr:	234
Title:	Cattle Weight Estimation from Dense Point Clouds
Authors:	Letícia Ferrari Castanheiro, Everton Castelão Tetila, Danielle Elis Garcia Furuya, João Paulo da Silva, Jayme Garcia Arnal Barbedo, Luciana Alvim Santos Romani and Edson Luis Bolfe
Abstract:	Cattle weight is essential for decision-making in precision livestock farming, directly supporting nutrition management, animal welfare, and production efficiency. Existing methods rely on close-range measurements or manual intervention, limiting scalability. This work proposes an workflow for cattle weight estimation based on point clouds derived from aerial images. RGB images acquired at low altitude were processed using Structure from Motion (SfM) techniques to generate dense point clouds. Individual animals were automatically segmented from the reconstructed 3D scene, and voxel-based volumetric features were extracted for each animal. Body weight was then estimated through linear regression models calibrated with ground truth measurements obtained from individual weighing. The proposed approach was evaluated on Nellore cattle in a feedlot environment and achieved a root mean square error (RMSE) of 8.35 kg, corresponding to an average relative error of approximately 2.29%. The results highlight the potential of UAV-based photogrammetry as a cost-effective decision support tool for digital and sustainable livestock management.
Download

Paper Nr:	240
Title:	A Physics-Informed Decision Support System for Airport Air Quality Management
Authors:	Abdellahi Babah, Souhila Arib, Hajer Baazaoui and Ayce Celikel
Abstract:	Nowadays, airport air quality management is increasingly requiring strategic decision-support systems able to integrate heterogeneous environmental data, operational constraints, and physical modeling principles into coherent analytical frameworks. Until now, approaches have either relied on regulatory dispersion models that provide interpretability but only limited adaptability or on purely data-driven models that offer predictive flexibility while lacking physical consistency and system-level integration. This paper presents a physics-informed digital twin decision support system for strategic airport air quality management. We introduce the Physics-Informed Graph-Voxel Fusion Transformer (Pi-GVFT) architecture that integrates three layers: a multi-resolution 3D voxel representation encoding the airport infrastructure, a dynamic physics-guided graph modeling evolving source-receptor interactions, and a probabilistic correction mechanism that enhances baseline simulations while preserving physical plausibility. We first present the technical foundations of the Pi-GVFT model, followed by an analysis of its theoretical constraints and practical applications. The proposed use cases demonstrate the feasibility and potential to transform airport environmental management, enabling a shift from a reactive to a proactive, data-driven approach.
Download

Paper Nr:	243
Title:	Robust Agentic AI: An Architectural Framework Using MCP and Deterministic Safety Guardrails in the Context of Mining in Industry 4.0
Authors:	Breno H. N. Andrade and Ricardo A. R. Oliveira
Abstract:	Mining automation faces a critical architectural challenge: how to deploy intelligent autonomous systems in extreme edge environments where cloud connectivity is unreliable and safety requirements are non-negotiable. This position paper introduces the Mine-Continuum Agentic Framework (MCAF), which embeds Small Language Models directly at the asset level while maintaining deterministic safety through formal verification. The framework addresses the latency-intelligence-safety trilemma-classical automation provides safety and low latency but lacks contextual intelligence, while cloud AI offers intelligence at the cost of unacceptable delays and connectivity dependencies. MCAF resolves this tension through a three-layer architecture combining edge-native reasoning with the Model Context Protocol for semantic grounding and LogicGuard for runtime enforcement via Linear Temporal Logic. Our comparative analysis shows how this approach bridges the semantic gap between probabilistic AI outputs and deterministic OT requirements-a gap existing paradigms leave unaddressed. The paper concludes with an open research agenda for Industry 5.0 human–machine collaboration in safety-critical industrial environments.

Paper Nr:	247
Title:	A Survey of LLM-Based Question Answering Architectures and the Role of Semantics in Enterprise Decision Support Systems
Authors:	Antony Seabra, Daniel Schwabe and Sergio Lifschitz
Abstract:	Decision support systems (DSS) are central to enterprise environments, where complex decisions require integrating heterogeneous information, applying domain criteria, and interpreting implicit knowledge. Recent advances in large language models (LLMs) offer new opportunities to support such processes through natural language interaction and flexible access to distributed data. However, current LLM-based question answering architectures remain limited in their ability to address decision-oriented queries, as they largely depend on retrievable information and lack explicit representations of domain semantics. This survey analyzes contemporary approaches-including retrieval-based, graph-based, and agentic architectures-with respect to their representational assumptions, reasoning mechanisms, and suitability for decision support. We identify a shared structural constraint across these paradigms: their reliance on accessible data rather than formalized conceptual knowledge. To address this limitation, we examine how semantic representations and structured knowledge models, such as ontologies and symbolic reasoning frameworks, can complement existing architectures. We conclude by outlining research directions for developing semantically grounded AI systems capable of supporting complex decision-making in enterprise contexts.
Download

Paper Nr:	260
Title:	A Sensor-Based Multi-Agent Ecosystem for Biomechanical Analysis: A Framework for Sensor Fusion, Digital Twins, and Autonomous Agent Reasoning
Authors:	Patrick B. N. Alvim, Jonathan C. F. da Silva, Vicente J. P. Amorim, Pedro S. O. Lazaroni, Mateus Coelho Silva and Ricardo A. R. Oliveira
Abstract:	Biomechanical analysis of human locomotion is an important tool in the orthopedic field and in sports. Current wearable systems perform indirect and low-fidelity measurements because they are made by single-point devices, such as smartwatches, which despite providing precise analysis, do not capture movement directly on the limb. To overcome this limitation, we present a direct capture framework using a wearable device composed of inertial sensors positioned on the lower limbs. This architecture sends the data to a Long Short-Term Memory (LSTM) neural network that initially classifies the type of activity into idle, walking, running, and uphill, and achieved an overall classification accuracy of 99%. Simultaneously, the data are sent to a mobile application reproducing the user’s movement in a real-time 3D virtual twin. Furthermore, we propose evolving this monitoring system into a proactive ecosystem based on multi-agent systems (MAS), integrating generative models, Large Language Models (LLMs), and Retrieval-Augmented Generation (RAG). With this architecture, we can distribute the processing of biomechanical data, specialized clinical reasoning, and user interaction among three autonomous agents, creating an active and predictive decision-support tool.
Download

Paper Nr:	261
Title:	The Impact of Artificial Intelligence on Organizational Innovation Initiatives: A Systematic Mapping Study
Authors:	Simone Santos, Alixandre Santana, Caio Gusmão, Diego Calixto, Eduardo Teles, José Nascimento, Júlia Fragoso, Júlio Bem and Vinícius Marçal
Abstract:	This study investigates how the adoption of Artificial Intelligence (AI) is transforming innovation initiatives within organizations, addressing the growing strategic role of AI in organizational processes and competitive dynamics. The study seeks to understand how AI is applied in innovation contexts, what benefits and challenges are reported, and how it influences organizational structures, processes, and decision-making. A Systematic Mapping Study was conducted following established guidelines, analyzing 72 studies from major scientific databases between 2020 and 2024, selected through predefined inclusion, exclusion, and quality criteria. The findings indicate that AI is increasingly integrated into innovation processes, enhancing data-driven decision-making, automation, and knowledge generation, while also revealing persistent challenges related to organizational culture, skills gaps, governance, and ethical concerns. The study highlights heterogeneous applications of AI across industries and innovation stages. In summary, the paper provides a structured overview of the current research landscape, identifies gaps in empirical and theoretical development, and offers implications for future research and managerial practice to better align AI capabilities with organizational innovation strategies.
Download

Paper Nr:	294
Title:	Multi-Sided Assembly Line Balancing Problem with Assignment Restrictions
Authors:	Abdolreza Roshani
Abstract:	Multi-Sided assembly line balancing problems commonly arise in plants producing large-sized, high-volume products such as automobiles. Unlike traditional assembly lines, multi-sided lines allow several parallel workplaces (sides) at each workstation, enabling operators or robots to perform different tasks simultaneously on the same product. This paper addresses the multi-sided assembly line balancing problem with assignment restrictions. In this problem, certain tasks must be assigned to the same side (e.g., tasks requiring the same tool or fixture to reduce tooling costs), while others cannot be assigned together (e.g., welding and painting tasks). These requirements are referred to as compatible (positive) and incompatible (negative) zoning constraints. A mixed-integer programming model is proposed to obtain optimal solutions by minimizing line length and the number of stations for a given cycle time. Computational experiments are conducted to evaluate the proposed modelling approach, and the results demonstrate its effectiveness and efficiency.
Download

Paper Nr:	297
Title:	Generative AI in Action: Use Cases, Challenges, and Enablers for SMEs in the High-Tech Manufacturing Industry in the Netherlands
Authors:	Evelyn Zuidema-Tempel, Ralf Kleijkers, Konrad Schneidenbach and Senem Yazici
Abstract:	Despite the growing interest in Generative AI (GAI), its application in small and medium-sized enterprises (SMEs) in the Dutch High-Tech Systems and Materials (HTSM) sector remains underexplored. This study empirically examines how SMEs adopt and apply GAI, and which technological, organisational, and environmental factors enable or constrain this process. Using the TOE framework, 11 semi-structured interviews were conducted with participants from seven SMEs in the eastern Netherlands. The findings show that GAI adoption is still largely exploratory. Most firms experiment with accessible tools for document analysis, content generation, and workflow support, while a small number of advanced adopters integrate GAI into broader business processes such as project initiation, contract evaluation, and production monitoring. The results further indicate that technological readiness alone is insufficient for successful adoption. Organisational conditions, including leadership support, internal champions, and opportunities for experimentation, play a central role in enabling firms to move beyond individual experimentation. In addition, external knowledge networks and collaboration with universities and innovation hubs support SMEs in accessing expertise and scaling applications. By linking these findings to the TOE framework, the study operationalises technological, organisational, and environmental factors through concrete GAI use cases, challenges, and enabling conditions observed in SMEs.
Download

Paper Nr:	304
Title:	COVADISE: Cover Based Anomaly Detection in Multi-Variate Time Series
Authors:	Pierre Lotte, André Peninou and Olivier Teste
Abstract:	Anomaly detection in Multivariate Time Series (MTS) is crucial for monitoring complex systems, where inter-variable dependencies encode essential behavioral patterns. Conventional approaches process all variables jointly, which can obscure localized relations and hinder interpretability in high-dimensional settings. We introduce COVADISE, a causality-preserving framework that splits variables into causally coherent subsets, locally trains instances of a classical anomaly detectors independently, and merges their outputs into a unified anomaly score. This design preserves causal structures while reducing dimensionality and enhancing model specialization. Experiments on synthetic and real-world datasets show that COVADISE significantly improves detection performance when causal relations are stable and remains competitive under noisy or drifting conditions. COVADISE thus provides a robust, model-agnostic approach for causality-aware anomaly detection in MTS.
Download

Paper Nr:	311
Title:	Impact of Feature Selection and Genetic Algorithms on Machine Learning Performance for Intraday Financial Trend Prediction
Authors:	Andrey V. Souza, Bruno L. Dalmazo, Giancarlo Lucca, Eduardo N. Borges, Richard F. Pinto, Fabian C. Cardoso, Viviane L. D. de Mattos and Rafael A. Berri
Abstract:	This study examines the effect of feature selection and Genetic Algorithm-based hyperparameter optimization on intraday financial trend forecasting in the Brazilian futures market. Using Mini Index (WIN) contracts from the BovDBv2 database, five feature selection methods were evaluated: Information Gain, ReliefF, PCA, Clas-sifierAttributeEval, and CorrelationAttributeEval. Random Forest and Multilayer Perceptron classifiers were employed, with optimal feature subsets selected through stratified cross-validation and model hyperparameters subsequently optimized by a Genetic Algorithm. All optimized Random Forest models outperformed the baseline accuracy of 0.6381, with PCA achieving the best RF result at 0.6555. The best overall performance was obtained by the MLP with PCA and GA-tuned hyperparameters, reaching 0.6571. These findings indicate that jointly improving feature representation and model configuration enhances generalization in noisy and nonstationary market prediction tasks.

Paper Nr:	314
Title:	Industrial Waste Reduction Applications of Artificial Intelligence: Progress, Challenges, and Future Directions
Authors:	Bilel Elayeb, Michael Lecointre and Maher Jridi
Abstract:	This survey reviews recent advances in applying Artificial Intelligence to industrial waste reduction. It highlights the use of machine learning, deep learning, predictive analytics, and digital technologies to improve waste classification, process optimization, resource efficiency, and circular economy integration. While AI has significantly enhanced operational efficiency and sustainability performance, key challenges remain, including data limitations, model interpretability, system integration, and implementation costs. The study outlines future directions such as explainable AI, real-time monitoring, hybrid modeling, and lifecycle-based optimization to support scalable and sustainable industrial transformation.
Download

Paper Nr:	317
Title:	Semantic Retrieval for the Reviewer Assignment Problem: An Evaluation of Transformer-Based Language Models
Authors:	Chara Tsirka, Vaios Stergiopoulos, Eleni Tousidou, Michael Vassilakopoulos and Antonio Corral
Abstract:	The Reviewer Assignment Problem (RAP) is a key component of peer-review systems, requiring accurate matching of submitted papers to suitable expert reviewers. As submission volumes grow, scalable and automated reviewer recommendation becomes increasingly necessary. In this work, we formulate RAP as a large-scale semantic retrieval task, where papers act as queries and candidate reviewers are ranked by semantic similarity between paper content and reviewer expertise. Datasets of papers and reviewer profiles are constructed from the AMiner citation network. Due to the lack of explicit reviewer assignment labels, we treat the authors of a paper as its most suitable reviewers. Two Sentence Transformer models, MiniLM and MPNet, are fine-tuned under multiple strategies and evaluated against the BM25 lexical retriever and the E5 dense retrieval model. Results show that transformer-based methods outperform lexical baselines, with fine-tuning further improving ranking performance. The best results are achieved by MiniLM using enriched pairs that include both authors and frequent citers as suitable reviewers. These findings demonstrate that fine-tuned transformer-based models provide an effective and scalable approach for automated reviewer assignment.
Download

Paper Nr:	319
Title:	Detection of Spoofing Attacks in Medical IoT Devices Based on the Adaptive Neuro-Fuzzy Inference System
Authors:	Andrew Mendonça Alaniz, Graçaliz Dimuro and Bruno L. Dalmazo
Abstract:	This paper investigates the detection of spoofing attacks in Internet of Medical Things (IoMT) networks, focusing on ARP spoofing as a critical threat capable of enabling impersonation and man-in-the-middle behavior in connected healthcare environments. Using a reproducible experimental pipeline on the CICIoMT2024 dataset, this study evaluates an Adaptive Neuro-Fuzzy Inference System (ANFIS) that combines low computational cost with rule-based interpretability, which is desirable in safety-critical scenarios. The results demonstrate that the proposed approach can effectively distinguish normal from spoofing traffic while maintaining a low rate of false alarms, suggesting practical potential for lightweight intrusion detection in clinical networks. Finally, the paper discusses paths for future work, including extending coverage to additional attack families, improving robustness under distribution shifts, and exploring more specialized versus more generalizable detection models.
Download

Paper Nr:	321
Title:	MOHICAN: Smart Cipher Algorithm for Secure Computational Face
Authors:	Joanna I. Olszewska
Abstract:	With the current growth of enterprise applications relying on facial identification, verification, or authentication, the secure storage of face pictures on devices such as mobile phones, on services such as the cloud or at the edge becomes of increasing importance. For this purpose, we propose a smart cryptographic system based on a Modified Hill Cipher Algorithm (MOHICAN) which allies automated facial detection with fast encryption of face images, whilst being efficient compared to state-of-art methods as demonstrated on standard image datasets.
Download

Paper Nr:	323
Title:	A Holistic Agent-Based Approach: Combining Data Architectures and Job-Shop Scheduling
Authors:	Marco Wölfel, Adrian Pfleiderer and Bernhard Bauer
Abstract:	Manufacturing systems must respond to volatile demand, heterogeneous data, and frequent disruptions while maintaining quality and throughput. This paper presents an agent-based framework that connects predictive analytics, job-shop scheduling, and data collection through a digital twin data architecture. At the core of the data architecture is a unified namespace structured according to the proposed data model, enabling continuous information exchange between shop floor execution, predictive analytics, and scheduling. The approach models machines, human operators, and software services as agents, with a central data access broker supporting the closed-loop feedback between execution data and scheduling decisions. This feedback loop allows historical and real-time process data to inform future scheduling decisions, parameter adaptation, and resilience strategies. The framework is designed for extensibility, domain independence, and variable modeling granularity, supporting both long-term planning and detailed execution control. A scenario-based evaluation in a collaborative manufacturing use case demonstrates how the architecture integrates operational technology, information technology, and predictive services to support adaptive, self-correcting, and resilient production.
Download

Paper Nr:	324
Title:	A Hybrid Multicriteria Decision Method Using AHP and PROMETHEE II for Serverless Provider Selection
Authors:	Kelwin Efrain Bagnhuk da Silva and Adriano Fiorese
Abstract:	The growing expansion and adoption of the serverless computing paradigm have driven the emergence of several new technologies and service configurations, making the issue of choosing the most suited serverless platform fitting technical and specific requirements increasingly challenging and complex for clients. Furthermore, it has become evident that the application of isolated multi-criteria analysis methods presents technological limitations and restrictions, mainly involving massive datasets, frequently resulting in a high computational overhead due to the quadratic growth of pairwise comparisons or high mathematical burden. To overcome these obstacles, this paper proposes and establishes a multi-layered hybrid method featuring three stages for the selection of serverless infrastructures. The method integrates an initial logical filtering layer based on Performance Indicators (PIs) and, as an intermediate layer, the Analytic Hierarchy Process (AHP), utilized solely to weight the analyzed criteria. Finally, the PROMETHEE II algorithm is used for the final evaluation. To validate the proposal, experiments assessing accuracy and computational time were executed using a dataset simulating potential provider alternatives. These results were compared against a baseline implementation and demonstrated that the hybrid strategy delivers adaptive computational performance intrinsically tied to the decision-maker’s constraint profile, achieving execution times up to 0.005s in restrictive scenarios, while sustaining near-perfect positional accuracy (100% in controlled scenarios, with rare outliers at large scale) up to 1,000 providers.
Download

Paper Nr:	47
Title:	Fairness-Conscious Feature Engineering and Clustering for Debt Recovery: Performance, Ethics, and Regulation
Authors:	Bianca Garcia Martins, Luiz Sergio de Souza and Solange N. Alves-Souza
Abstract:	Customer segmentation is critical in debt recovery, enabling strategies that balance efficiency and retention. However, automated segmentation can inadvertently perpetuate biases, violating regulations like the EU AI Act, GDPR, and LGPD. This paper investigates trade-offs between clustering quality and algorithmic fairness in unsupervised learning. A reproducible PySpark pipeline using the Medallion Architecture is proposed to compare three feature engineering scenarios: (1) baseline; (2) “fairness through unawareness”; and (3) a “fairness-conscious” approach involving proxy removal and linear debiasing. Using the Default of Credit Card Clients dataset from the UCI Repository, containing 30,000 records from Taiwan, a grid search was performed (k ∈ {3,4,5}). The fairness-conscious scenario achieved superior clustering quality (Silhouette +43.5%, Calinski-Harabasz +202.7%, WSSE -67.0% vs. baseline) while satisfying the 80% Rule for gender (DP Ratio > 0.8). However, age-related disparities persisted (DP Ratio < 0.8), highlighting indirect bias through behavioral proxies. The findings demonstrate that careful feature engineering can simultaneously improve clustering quality and mitigate bias. Practical guidelines for auditing and mitigating bias in financial segmentation are provided.
Download

Paper Nr:	48
Title:	Design and Evaluation of a Modular Multilingual RAG Architecture for Trustworthy and Auditable AI Systems: The PrivacyEthical Chat Tool
Authors:	Edna Dias Canedo, Stefano Luppi Spósito, Flávio Garcia Praciano, Thiago L.de Sousa, Paulo Henrique B. Rodrigues, Geraldo Pereira Rocha Filho and Fábio L. L.de Mendonça
Abstract:	The growing adoption of Large Language Models (LLMs) in institutional settings raises critical challenges regarding trustworthiness, transparency, and ethical compliance. In domains such as ethics, privacy, and information security, automated systems must provide accurate and traceable responses. Retrieval-Augmented Generation (RAG) addresses these needs by grounding outputs in external, verifiable sources, yet little research has examined how to design and evaluate RAG for multilingual, governance-sensitive environments. We investigate how a modular RAG architecture can be engineered to produce high-precision, auditable outputs in multilingual institutional contexts by combining heuristic language detection, domain-specific routing, and semantic retrieval. We implemented a three-stage retrieval pipeline: (i) heuristic language detection, (ii) domain routing, and (iii) semantic retrieval with high-precision vector embeddings and evaluated it on a bilingual (Portuguese/English) dataset of ethical and privacy requirements. Language detection achieved 84.3% accuracy; embeddings exhibited clear semantic separation between domains; and cosine similarity for relevant documents consistently exceeded 0.7, demonstrating retrieval robustness. The system also generated structured, traceable user stories grounded in validated ethical and privacy requirements. We contribute a validated multilingual RAG architecture that operationalizes trustworthy and explainable AI via transparent retrieval and accountable response generation. The design supports institutional governance by combining accuracy, interpretability, and scalability, advancing ethical, auditable, and reliable AI across organizational contexts.
Download

Paper Nr:	63
Title:	Towards Characterizing Crystal Structures with CNNs
Authors:	Rodrigo Guedes de Souza, Luiza Castilho Ereno, Analúcia Schiaffino Morales, Antonio Carlos Sobieranski, Fabricio Ourique and Alison R. Panisson
Abstract:	Crystal characterization is a critical task in chemistry and pharmacy, with direct implications for drug development and quality assurance of healthcare products. Despite its importance, the process remains largely labor-intensive and reliant on expert analysis. This study investigates the automation of crystal characterization using Convolutional Neural Networks (CNNs) to classify crystal shape and density. Two modeling strategies are explored: a multi-stage classification approach, achieving accuracies of 96.25% for density and 82.50% for shape, and a single-stage model, reaching an overall accuracy of 92.50%. The results demonstrate the potential of deep learning techniques to streamline crystal analysis, improving efficiency and consistency in both research and production environments.
Download

Paper Nr:	84
Title:	Post-Session LLM-Based Analysis of Multi-Role Teleconsultations for Audit and Medical Education
Authors:	Santiago Sánchez-Sobrino, David Vallejo, Francisco M. García, Sergio Martínez-Cid and Javier A. Albusac
Abstract:	Synchronous video teleconsultations between primary care staff and hospital specialists can provide timely expert input without moving the patient, yet the resulting recordings are often left as opaque audiovisual files that are costly to review and difficult to reuse for audit or teaching. This paper presents a post-session analysis pipeline that turns each multi-role consultation into a repository-ready case fiche. The approach combines speech processing (time-aligned transcription and speaker/role attribution) with LLM-based temporal structuring, structured extraction, and section-level summarisation. A key design goal involves that every extracted item and summary statement is linked to time references over the transcript and the original audiovisual stream, so reviewers can quickly verify evidence and navigate to relevant moments. The work is situated within the MRP-5G system context, but focuses specifically on the post-session analytics layer and its repository-oriented outputs. We also describe a reproducible evaluation protocol based on a synthetic multi-role corpus generated by our web prototype, providing controlled ground truth (scripts, turn boundaries, roles, and timestamps) and enabling largely automatic metrics for role attribution, traceability, and retrieval/navigation. Finally, a qualitative spot-check illustrates typical failure modes and practical mitigations when producing structured, evidence-linked artefacts.
Download

Paper Nr:	96
Title:	Towards an Assistant for Legal Framing in Agricultural Oversight: A RAG and LLM Chain Approach
Authors:	Fernando Soso Girardi, Thiago Paulo Both, Pedro Bilar Montero, Gabriel Rodrigues da Silva, Francesco Krum, Lauren Nicoloso Casarin, Alencar Machado and Vinícius Maran
Abstract:	The management of sanctioning proceedings in agricultural defense agencies faces challenges regarding efficiency and legal certainty, often stemming from the complexity of interpreting sanitary regulations. This study investigates the development of a Decision Support System for the Official Veterinary Service of Rio Grande do Sul (SEAPI-RS), designed to assist in the legal classification of Infraction Notices. The methodology proposes a Multi-Agent System architecture based on Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG), employing a sequential pipeline of sanitization, profile identification, and legal analysis. Validation, conducted via mixed methods on a controlled historical dataset (N=12) and operational field tests (N=8), demonstrated high accuracy in actor recall within the evaluated sample but revealed limitations in realworld scenarios, specifically regarding sensitivity to input quality and semantic retrieval biases. The study concludes that, while feasible, deploying LLMs in compliance tasks requires evolving towards Bottom-Up architectures supported by intelligent indexing to ensure a robust distinction between duties and penalties.
Download

Paper Nr:	111
Title:	TrendBot: A Machine Learning Framework for Financial Asset Trend Prediction Using Time-Series Sliding Windows
Authors:	Fernanda Bertão Mota and Mateus Coelho Silva
Abstract:	Financial market prediction remains a challenging task due to the high volatility and noisy nature of price time series. In recent years, machine learning and deep learning techniques have been increasingly explored as tools to model short-term price movements and support trading decisions. However, many proposed approaches rely on complex architectures, which may limit interpretability and practical applicability. This study addresses the problem of short-term stock price direction prediction using a simplified deep learning framework. We investigate whether a fully connected deep neural network, combined with a structured sliding window representation of historical prices, can effectively capture relevant temporal patterns without relying on recurrent or hybrid architectures. Here, we propose a classification-based approach in which historical price information is transformed into fixed-length input vectors using a 20-day sliding window. The model predicts whether the asset price will increase on the following trading day. The network is trained using normalized price variations and incorporates regularization techniques and class weighting to address data imbalance. The experimental results show that, while overall accuracy remains close to baseline levels commonly observed in financial forecasting tasks, the probabilistic output of the model provides valuable information for decision-making. By applying confidence thresholds to the softmax output, it is possible to improve precision and reduce falsepositive trading signals significantly.
Download

Paper Nr:	113
Title:	Neighborhood-Search-Based Path Planning Algorithm Applied to a Python Robotic Simulation Environment with Collision Memory
Authors:	Polliana Barelli Trentini and Mateus Coelho Silva
Abstract:	Path planning is a critical field for mobile robots, responsible for guiding robots from an initial point to a goal point with a collision-free, smooth, and fast trajectory. Path planning algorithms ensure the creation of these trajectories for robots, while robot simulators help researchers experiment with algorithms without using actual hardware, making these tests safer and cheaper. In this context, our work focuses on developing a neighborhood-search algorithm with memory collision, applied to a lightweight Python simulator. We want to show that we can reach good results even when using a simple algorithm. Since we applied our algorithm to a simulator, we had the freedom to test on two different maps, divided into low and high obstacle density. To make those tests, we used different parameters of our algorithm, exploring the maps in distinctive ways and finding successful paths while also evaluating its failures in a few situations. Our results indicate that our algorithm is a promising work with a feasible technique, since it was capable of finding trajectories in both maps.
Download

Paper Nr:	119
Title:	Stigmergic Navigation Using Reactive Potential Fields and Ant Colony Optimization
Authors:	Mayara Gonçalves de Siqueira and Mateus Coelho Silva
Abstract:	Autonomous navigation of mobile robots in unknown environments with obstacles is a fundamental challenge in robotics. Although the isolated use of Artificial Potential Fields (APF) ensures safe reactivity, this technique often results in inefficient trajectories. The central problem is to overcome these limitations by combining immediate safety with route optimization. In this work, we propose a hybrid stigmergic approach in which a Leader robot (APF) traces a safe digital pheromone path. In contrast, a Follower robot uses Ant Colony Optimization (ACO) to optimize the route on this map. The simulation results demonstrate that the Follower not only maintains safety but also optimizes the path length, smoothing curves, and tangenting the Leader’s path. This architecture demonstrates the efficiency of combining APF and ACO, offering a robust solution for autonomous navigation.
Download

Paper Nr:	124
Title:	Android Malware Detection through Static Analysis Using Machine Learning
Authors:	Gabriel Leão, Edward David Moreno and Methanias Colaço Júnior
Abstract:	Android is currently the most widespread mobile operating system, making it a primary target for malware that can compromise the security and privacy of millions of users. This work investigates the effectiveness of machine learning algorithms in performing Android malware detection via static analysis. We evaluate seven models - Logistic Regression, Random Forest, Gradient Boosting, XGBoost, SVM, KNN, and MLP Neural Network - using different subsets of features (API Calls, Intents, Permissions, and their combination). We additionally assess the impact of class imbalance by applying SMOTE and the effect of dimensionality reduction via PCA. Experiments are conducted on the Drebin dataset, which contains 5,560 malware samples and 9,476 benign applications. Our results show that combining all feature types yields superior performance, that ensemble models achieve higher accuracy, and that class balancing significantly improves recall without substantially degrading precision.
Download

Paper Nr:	148
Title:	Mapping Scientific Output to the Sustainable Development Goals (SDGs) via Multilingual Multi-Label Classification
Authors:	Vinícius Augusto Alves Santos Mello, Carlos Basílio Pinheiro and Wladmir Cardoso Brandão
Abstract:	The United Nations 2030 Agenda defines 17 Sustainable Development Goals (SDGs) that increasingly guide how universities demonstrate the societal relevance of their research. However, SDG attribution in scientific publications still often relies on keyword searches. This study presents a reproducible pipeline for assigning SDGs to publications through multilingual multi-label text classification, using as case study the journal articles listed in the Lattes CVs of faculty members at the Universidade Federal de Minas Gerais (UFMG). We compare five supervised models-linear SVM, XGBoost, Random Forest, XLM-RoBERTa, and DistilBERT-trained on Elsevier SDG labels and externally validated on the OSDG Community Dataset. Transformer-based models outperform TF–IDF baselines in both micro- and macro-F1, highlighting the importance of contextual representations for SDG mapping. Based on these results, XLM-RoBERTa is selected for inference on UFMG’s scientific output, revealing a research profile dominated by SDG 3 and SDG 7. The proposed pipeline offers a scalable and reproducible approach to support institutional planning and SDG-oriented transparency.
Download

Paper Nr:	152
Title:	PRIORI: An Ontology-Based Approach with a Context-Sensitive Chatbot for Classifying Risk and Protective Factors in Child Health
Authors:	Antônio Silva, Andrey Pimentel, Tatiana Riechi and Thaís Fogaça
Abstract:	This article presents PRIORI, an integrated proposal aimed at supporting the structured collection of information and the classification of risk and protection in child care services. The approach combines a modular architecture based on chatbot, ontology, and a rule-based symbolic inference mechanism, aiming to ensure semantic consistency and traceability in the interpretation of the collected data. As main contributions, the following stand out: an architecture that integrates interaction and communicational adaptation modules (ADAPTA) with an ontological base and an inference engine; the preliminary ontological modeling of the social domain of the child risk and protection inventory, organized to favor interpretability and inspection; and an initial prototype that demonstrates dialogue strategies adapted to different user profiles. The results highlight the potential of the approach to support the structured collection and interpretation of biopsychosocial factors in child care contexts, offering a conceptual and technical basis for applications in real-world scenarios.
Download

Paper Nr:	204
Title:	A2-ANFIS: Exploring Triangular Norms for Antecedent Generation in Adaptive Neuro-Fuzzy Inference Systems
Authors:	Bruna Novack, Juliano Buss, Gabriel Silva, Helida Santos, Giancarlo Lucca and Renata Reiser
Abstract:	Neuro-Fuzzy systems, such as the Adaptive Neuro-Fuzzy Inference System (ANFIS), are essential for modeling complex non-linear relationships by combining the interpretability of fuzzy logic with the learning capabilities of neural networks. However, the standard ANFIS architecture is typically limited to the Product t-norm (TP) in its antecedent layer, which may not capture all nuances of uncertainty in diverse datasets. This position paper proposes the A2-ANFIS architecture, a generalization that introduces an Aggregator Selection Module to replace the fixed TP with various classes of triangular norms, including non-parametric and parametric variants. We evaluated the proposed model employing fifteen benchmark datasets and twelve different t-norms. Experimental results, obtained via cross-validation, demonstrate that A2-ANFIS outperformed the baseline in 73.3% of the cases, with the Sugeno-Weber and Schweizer-Sklar operators showing significant consistency. These findings suggest that the flexibility in choosing the antecedent aggregator is a key factor in improving predictive accuracy in neuro-fuzzy modeling.
Download

Paper Nr:	212
Title:	Prediction Sets as a Decision-Support Interface for Critical Classification Tasks
Authors:	Morteza Mohammady Gharasuie, Fengjiao Wang and Ravi Mukkamala
Abstract:	Machine learning classifiers are increasingly used in high-stakes decision-making systems, yet their outputs are typically restricted to a single predicted label based on maximum estimated probability. While effective in many settings, this practice can obscure uncertainty and lead to overconfident decisions, particularly under extreme class imbalance like long-tailed data distributions, where rare classes are more error-prone. This position paper argues that prediction sets offer a more appropriate inference-time interface for critical classification tasks by returning a set of plausible labels with statistical guarantees on true-label inclusion. We frame prediction sets as a model-agnostic, safety-oriented mechanism that better supports human-in-the-loop decision-making and risk-aware deployment. Minimal evidence on a long-tailed benchmark supports this idea. For the rarest class group, Top-1 accuracy is relatively low, ranging from 0.6843 to 0.7847 across training methods. Prediction-set inference, however, improves true-label inclusion to 0.8356–0.9583 with standard conformal prediction sets and to 0.96–1.0000 with newer methods. These findings suggest that prediction sets can offer more reliable support for rare and ambiguous cases.

Paper Nr:	216
Title:	An AI-Based Tool for Training Facial Emotion Recognition and Production in Patients with Schizophrenia: Development and Preliminary Results
Authors:	Marcel Antunes Raposo, Luciano Silva, Luciana Amorim and Roberto Moraes Cruz
Abstract:	Schizophrenia is a psychiatric disorder characterized by cognitive and social impairments, including difficulties in recognizing and expressing facial emotions. Considering the critical role of social cognition in the functioning and quality of life of individuals with schizophrenia, a web-based computational tool grounded in artificial intelligence and affective computing techniques was developed and preliminarily evaluated to train the recognition and production of facial emotional expressions. The system integrates automated facial expression analysis to provide objective performance scores, supporting structured emotional skills training. The instrument was developed following usability principles suitable for the target audience and employs images and audio stimuli to promote engagement. It was evaluated with four patients receiving specialized outpatient care using structured tests. Additionally, specialist clinicians conducted an independent assessment. The results indicated higher accuracy in recognizing positive emotions and provided preliminary feasibility indications of the tool as a complementary resource for AI-assisted social cognition training. Qualitative analysis provided indications of favourable user acceptance and preliminary expert validation. Despite the limited number of participants, the findings provide preliminary exploratory indications of the tool’s potential and support recommendations for refinement and broader evaluation in future studies. This study does not aim to demonstrate clinical efficacy, but rather to explore feasibility and interaction patterns in a pilot setting.
Download

Paper Nr:	228
Title:	How Well Do LLMs Pass Brazilian Exams? Evaluating AI Performance on the POSCOMP and OAB Bar Examination
Authors:	Matheus Lopes Raposo Lima, Lisandra de Melo Lima, Hellen Guterres França, Anselmo Cardoso de Paiva, Lisle Faray de Paiva and Jan Egger
Abstract:	Large Language Models (LLMs) can support students preparing for competitive entrance examinations, yet their effectiveness in domain-specific, non-English, and regionally contextualized assessments remains underexplored. This study compares seven leading LLMs - Claude Sonnet 4.5, ChatGPT-5, DeepSeek-V3.2, Gemini 3 Pro, Qwen3, LLaMA 4 Maverick, and Mistral - on two major Brazilian standardized exams: POSCOMP (National Graduate Entrance Exam in Computing) and the first phase of the Brazilian Bar Examination (OAB). Questions were extracted from official PDF documents, processed through automated pipelines, and submitted to each model via API. Performance was evaluated using accuracy, BERTScore, and ROUGE metrics against official answer keys. Results reveal substantial cross-domain variation. Claude achieved the highest accuracy on the OAB, followed by Qwen and LLaMA, while Gemini and Mistral showed lower performance. In POSCOMP, Gemini performed best, followed by DeepSeek, Claude, and Qwen, whereas Mistral obtained the lowest score. Claude also achieved consistently high BERTScore and ROUGE-L values, indicating balanced performance across domains. Overall, findings demonstrate that no single model dominates across tasks, highlighting the importance of informed model selection for AI-assisted exam preparation in regionally grounded academic contexts.

Paper Nr:	230
Title:	Leveraging Topic Modeling for Identifying Individual Expertise in Knowledge Management
Authors:	Ekaterina Mashina and Florian Wahl
Abstract:	In this article, we present methods for clustering expert-authored documents into thematic groups. We introduce an interpretable LDA-based pipeline that quantifies individual expertise and generates knowledge maps for knowledge-management systems. Applied to a corpus of 251 research articles, the pipeline achieved a coherence score of Cv = 0.57 and demonstrated strong alignment between statistical and semantic structures (ρ = 0.83). The results of the digital experiment indicate that Latent Dirichlet Allocation can be used for semantic text analysis to identify latent topics and distinguish them effectively, reflecting an employee’s expertise in domains associated with specific terminology. We further analyze factors supporting the integration of topic modeling as a foundation for textual analysis in the development of a unified knowledge-management system. The presented methods provide a basis for designing practical solutions to identify and formalize knowledge for subsequent integration into a company’s consolidated knowledge base.
Download

Paper Nr:	235
Title:	Transformer Attention Modifications and Pre-Aggregation Functions: A Systematic Review
Authors:	Diego Duarte Bottero, Joelson Sartori Junior, João Pedro Stone Moreira, Lizandro de Souza Oliveira, Giancarlo Lucca, Bruno L. Dalmazo and Rafael Alceste Berri
Abstract:	Sentiment analysis is a fundamental task in Natural Language Processing (NLP), widely applied to understanding opinions and emotions in domains such as e-commerce, social media, and decision support systems. This systematic review analyzes recent advances in aggregation functions and architectural modifications proposed for Transformer-based models, as well as their adoption in sentiment analysis tasks. Results show that aggregation mechanisms and architectural innovations are concentrated in non-textual domains such as computer vision and remote sensing. In contrast, sentiment analysis research remains dominated by standard pretrained models, primarily BERT, RoBERTa, DistilBERT, and ELECTRA, focused on fine-tuning and optimization rather than structural changes. This review identifies a clear gap between architectural innovation and practical adoption in NLP, highlighting opportunities for integrating advanced attention mechanisms into Transformer-based sentiment analysis.
Download

Paper Nr:	257
Title:	A Systematic Literature Review on Classical Data Encoding Strategies for Hybrid Quantum Machine Learning
Authors:	Cecilia Botelho, Gabriel Rosa de Oliveira Silva, Eduardo Timm Buss, Giancarlo Lucca, Helida Santos, Adenauer Yamin, Renata Reiser and Lizandro S. Oliveira
Abstract:	Quantum Machine Learning (QML) has gained attention through hybrid architectures that combine classical optimization with quantum models such as Variational Quantum Circuits (VQCs) and Quantum Neural Networks (QNNs). In these approaches, the encoding of classical data into quantum states plays a central role, influencing expressiveness, trainability, and quantum resource usage under Noisy Intermediate-Scale Quantum (NISQ) constraints. Despite the growing number of proposed encoding strategies, their evaluation remains fragmented and lacks consistent reporting criteria. This paper presents a Systematic Literature Review of classical data encoding strategies adopted in hybrid QML models. Following the PRISMA 2020 guidelines, studies published between 2020 and 2025 were identified and assessed using predefined inclusion, exclusion, and quality criteria. Fourteen primary studies were selected for analysis. The review examines the types of encoding strategies employed, the evaluation metrics reported, and how quantum resource requirements and feasibility on NISQ devices are addressed. The results reveal a predominance of rotation-based feature maps and variational encodings, while systematic comparisons and resource-aware evaluations remain limited. This synthesis highlights the need for more standardized evaluation practices in QML data encoding.
Download

Paper Nr:	278
Title:	Fusion of Visual Signal Representations via Multi-Head Attention for Spectrogram Classification
Authors:	Caio Grasso, Pedro Guedes, Pedro Coelho, José Franco Amaral, Thiago Carvalho and Michel Tcheou
Abstract:	This work investigates the use of multi-head attention for fusing visual representations of acoustic signals in the context of spectrogram classification. The central idea is to model different spectrogram pre-processings as complementary views and to learn an adaptive fusion through a self-attention block. As a case study, we consider a real underwater acoustic dataset composed of three classes of grayscale spectrograms. Each sample is represented by three filtered views (original, smoothed, and edge-enhanced), processed by a shared ResNet-18 backbone and fused using a multi-head attention module. We compare the proposed approach with a strong single-view CNN baseline and classical machine learning methods. Experimental results show that while multi-view fusion improves performance for weaker representations (STFT), it does not outperform a strong single-view CNN on Mel spectrograms. These findings highlight both the potential and the limitations of attention-based multi-view fusion under strong baselines.
Download

Paper Nr:	279
Title:	GenAI-Content Explainable Detection System
Authors:	Lewis Moore and Joanna Olszewska
Abstract:	With the rise of Generative Artificial Intelligence (GenAI) and especially Large Language Models (LLMs) to produce essays, courseworks, reviews, papers, books, news, etc., it is important to be able to differentiate between such AI-generated content and human-generated one. Hence, our paper proposes to detect text which has been produced by GenAI-based chatbots. Our detection system uses an explainable Artificial Intelligence (XAI) approach based on a Multinomial Naive Bayes (MNB) model, and it offers a free, open-source, and data-privacy compliant tool for enterprises. Furthermore, it shows promising performances in terms of accuracy and sustainability when compared with existing freewares.
Download

Area 3 - Information Systems Analysis and Specification

Full Papers

Paper Nr:	56
Title:	Automated Code Review with LLMs: A Comparative Study of Recent Models in CI/CD Pipelines
Authors:	Thiago Paulo Both, Diones de Vargas Dutra, Pedro Bilar Montero, Fernando Soso Girardi, Francesco Krum, Gabriel Vieira Casanova, Alencar Machado and Vinícius Maran
Abstract:	Manual code review is an indispensable practice to ensure software quality, but it represents a significant bottleneck in modern development pipelines due to its time-consuming and inconsistent nature. To address these challenges, this paper explores the automation of code review through the application of Large Language Models (LLMs). We present a practical, comparative study of two advanced models, Gemini-2.5-Pro and Gemma-3-27b-it, integrated directly into a real-world Continuous Integration/Continuous Delivery (CI/CD) pipeline. The proposed solution was validated within the development environment of the Animal Health Defense Platform of Rio Grande do Sul (PDSA-RS), a strategic system of high economic and sanitary importance. The central objective is to evaluate the feasibility and efficiency of these LLMs in an automated code review process, offering insights into their potential to enhance the software engineering workflow, reduce developer overhead, and improve code quality in a critical application.
Download

Paper Nr:	65
Title:	ODMD: An Ontology-Driven Method for Metapath Design in Heterogeneous Graphs
Authors:	Silvio Fernando Angonese and Renata Galante
Abstract:	Heterogeneous graphs are used in Information Systems to represent complex domains involving multiple entity types, multimodal content, and higher-order semantic relations. Metapaths, typed sequences of relations that encode semantic paths between heterogeneous entities, play a central role in capturing semantics. They are often defined manually, guided by experts, intuition, or trial-and-error. This lack of methodological support limits reproducibility, semantic coherence, and reuse of metapaths across domains. To address this gap, this paper proposes ODMD – Ontology-Driven Method for Metapath Design, a systematic method for deriving, validating, and selecting metapaths from an explicit domain ontology. ODMD combines ontological constraints, competency-based filtering, and lightweight predictive scoring to generate metapaths that are semantically valid and empirically useful. The method integrates into the AGHE - Approach for Generating Enhanced Heterogeneous Node Embeddings in Heterogeneous Graphs pipeline, enabling metapath-enriched heterogeneous node embeddings composed of Features and Metapath-based encodings. Experimental study on an authorial Person-Relationships graph shows that metapaths generated by ODMD outperform ad hoc alternatives. Results indicate improved classification effectiveness, reduced variability, and coherent latent spaces, confirming that ontological guidance enhances the quality and stability of heterogeneous embedding compositions. These findings highlight the value of ODMD for Information Systems scenarios requiring semantic consistency and trustworthy representations.
Download

Paper Nr:	144
Title:	Human Factors in Agile Methodologies: A Survey Analysis
Authors:	Raquel C. Martins, José L. F. da Silva Junior and Michel S. Soares
Abstract:	Agile methodologies have been widely adopted in software development over the past two decades. However, their impact on the work environment is often overlooked. This paper investigates agile methodologies from the perspective of the human factor, focusing on their physical, emotional, and organizational impacts in the work environment. The study was conducted through a survey with technology professionals working in contexts with and without the adoption of agile practices. The results indicate that agile methodologies are perceived as effective in improving communication, productivity, and delivery quality. At the same time, relevant impacts emerge related to psychological pressure, mental fatigue, and the need for constant availability. Factors such as psychological safety, collaboration, and leadership support are shown to be central to professionals’ experiences. The findings suggest that the benefits of agile methodologies strongly depend on the organizational context and on how they are implemented. It is concluded that a more human-centered and sustainable approach is necessary to enhance organizational gains and reduce individual costs.
Download

Paper Nr:	149
Title:	A Systematic Mapping of Large Language Models for Improving Natural Language Software Requirements
Authors:	Raiane Eunice C. Santos Fernandes, Fábio Gomes Rocha and Rheidner Achiles Couto S. Fernandes
Abstract:	The advancement of Large Language Models (LLMs) has impacted Requirements Engineering, an area traditionally challenged by the evaluation of the quality of requirements expressed in natural language. Despite the growing interest from the scientific community, there are still gaps in understanding how these models have been used to systematically support the evaluation and improvement of requirements quality. In this context, this study presents a systematic mapping aimed at investigating the application of LLMs in the evaluation, refinement, and improvement of textual requirements, considering aspects such as usage contexts, techniques adopted, tools employed, requirements quality metrics, and reported empirical evidence. The mapping was conducted based on the PICOC principle, using searches performed in four digital databases. The results indicate a predominance of the use of proprietary LLM models and suggest that these technologies have the potential to improve quality attributes such as clarity, completeness, and consistency. However, challenges persist regarding the reliability of the results and practical adoption, since most studies were conducted in academic contexts, with limited validation in real-world industrial settings.
Download

Paper Nr:	165
Title:	Hybrid Prioritization of Self-Admitted Technical Debt and Static Code Analysis: A Multiple-Case Study
Authors:	Rafael Santos Silva, Rodrigo Rebouças de Almeida and Wylliams Barbosa Santos
Abstract:	Technical Debt (TD) management remains a challenge in software projects, especially when priorities conflict across technical and business aspects. Debts identified through Static Code Analysis (SCA) provide objective information, while Self-Admitted Technical Debt (SATD) reflects the team’s subjective perception of technical and business concerns. However, when prioritized in isolation, neither approach provides a comprehensive view, often resulting in prioritization decisions with limited impact or misalignment with project goals. This study aims to understand how software professionals prioritize technical debts identified by SCA and SATD when evaluated collaboratively, as well as to identify the criteria used in hybrid prioritization. A multiple-case study was conducted with technical and business experts across real-world projects. Each project included debts identified through SCA and SATD, which were first prioritized individually and then collaboratively reviewed to produce a unified prioritized list. The findings indicate that hybrid prioritization is a negotiation-based process that balances technical characteristics with factors influencing business impact. The study demonstrates that hybrid prioritization fosters a shared understanding between technical and business experts, promoting decision-making that is both balanced across stakeholder perspectives and sustainable for long-term software maintenance.
Download

Paper Nr:	190
Title:	Parallelism Techniques with GPUs for Viral Variant Analysis
Authors:	Gabriel Leoni Duarte, Matheus Carreira Andrade, Clarisse Midori Yoshimura Torres, Gilberto Vaughan, Raghuvir Krishnaswamy Arni, Vitoria Zanon Gomes and Geraldo Francisco Donegá Zafalon
Abstract:	Bioinformatics increasingly depends on high-performance computing to keep pace with next-generation sequencing pipelines, where hundreds of thousands, in some case billions, of aligned genomic sequences must be compared and summarized efficiently, to allow the researchers to make inferences and to take some decisions over the next analyzes. One of the most benefited of this operation is viral variant analysis, which is a field of bioinformatics where it is possibile to understand the behavior of viral evolution, specially in humans, and the patients' response to new drugs and treatments. Thus, this paper presents a GPU-accelerated approach for large-scale analysis of viral variants based on massive Hamming-distance computations and the construction of One-Step Networks under a fixed distance threshold. Experiments were performed on two distinct CPU/GPU environments using an identical measurement protocol, reporting execution time, throughput (sequences per millisecond), and speedup. The results show consistent GPU advantages over multi-core CPU execution, with speedups between 1.5 and 1.7 times and higher throughput across datasets, supporting practical analysis on a local workstation without requiring clusters. In addition, the One-Step Networks produced from different timepoints reveal clear connectivity patterns that complement quantitative performance evaluation.
Download

Paper Nr:	198
Title:	A Sentiment Analysis Study Based on LLMs of an Open Source Software Project
Authors:	Rafael Takeguma Goto, Helena Carvalho Leal, Methanias Colaço Júnior and Glauco de Figueiredo Carneiro
Abstract:	Open-source software (OSS) development depends heavily on asynchronous, text-based communication. While the sentiment within these discussions is a critical indicator of team dynamics and project health, widely used platforms lack analysis tools tailored for software engineering (SE) contexts. This study leverages recent advancements in Large Language Models (LLMs) to classify sentiment in issue comments from the highly active Grafana project. We employ a zero-shot and few-shot prompting strategy that integrates issue context with exemplar annotations to evaluate six state-of-the-art models: GPT-4o mini, GPT-4.1 nano, GPT-5 nano, Gemini 2.0 Flash, Gemini 2.5 Flash, and Gemini 2.5 Pro. Our findings demonstrate that LLMs are highly effective for SE sentiment analysis (SA), though they exhibit distinct model-specific biases. Finally, we provide practical recommendations for integrating sentiment insights into project management workflows.
Download

Paper Nr:	253
Title:	Does Gamification in Requirements Engineering Work? An Updated and Extended Systematic Literature Review
Authors:	Eduardo Bulling Couto, Gilleanes T. A. Guedes, Paulo Silas Severo de Souza, Ricardo Vilela, Pedro Henrique Valle, Ana Carolina Oran and Williamson Silva
Abstract:	Requirements Engineering (RE) activities are key to the success of software development projects. As software systems grow in complexity and stakeholder needs continue to evolve, innovative approaches are needed to improve the efficiency and effectiveness of RE processes. Gamification has emerged as a promising strategy to boost engagement and improve outcomes in RE. It is defined as the use of game-design elements in non-game contexts (e.g., challenges, rewards, and competition). It has the potential to increase participant motivation, making the elicitation, analysis, and documentation of requirements not only more enjoyable but also more effective. In this study, we explore the application of gamification in RE processes, providing a comprehensive overview of the gamification elements and techniques employed, the specific RE subprocesses addressed, and the resulting impacts. To achieve this, we updated and extended a previous Systematic Literature Review (SLR) on gamification and its application to RE activities to incorporate new evidence and provide a more current and comprehensive synthesis of the field. Our findings indicate the use of 74 distinct gamification elements in RE, with points, leaderboards, and badges emerging as the most frequently reported, mainly in elicitation, analysis, and prioritization activities. The results suggest that gamification may enhance stakeholder engagement, motivation, collaboration, and communication, leading to higher-quality requirements and more efficient RE processes. The review also highlights challenges associated with gamification, such as design complexity, potential distractions, and the sustainability of game elements in long-term projects. This study offers valuable insights for RE professionals and researchers by providing practical guidance on how to implement gamification to improve the efficiency and quality of requirements-related activities.
Download

Paper Nr:	282
Title:	RouteBastion: Proposal and Performance Evaluation of a Unified API Broker for Vehicle Routing Problems
Authors:	Pietro Piva Vieira, Bruno L. Dalmazo and Rodrigo Brandão Mansilha
Abstract:	The growing use of cloud services for routing problems has made issues like being tied to one provider and dealing with different APIs more challenging. This paper presents RouteBastion, a unified API broker that abstracts multiple VRP providers through a single scalable and extensible interface, enabling dynamic provider selection based on configurable criteria. The architecture was qualitatively evaluated using ATAM and ALMA, and empirically assessed through controlled load experiments with k6 in a containerized environment. Results show near-linear scalability up to 300 concurrent users and stable latency under both constrained and unconstrained conditions.
Download

Short Papers

Paper Nr:	30
Title:	Software Architecting in the Brazilian Industry: Results of a Prior Survey
Authors:	Valdemar Vicente Graciano Neto, Mohamad Kassab, Diana Lorena Santos, Andrey Gonçalves França, Edson OliveiraJr, Rafael Z. Frantz, Ahmad Mohsin and Marcos Kalinowski
Abstract:	Context: Software architecture is essential for system quality and long-term sustainability, yet little is known about how the architect role is structured in Brazilian companies. Objective: This study examines how Brazilian organizations incorporate software architecture into their development processes, focusing on the presence and formalization of software architects, and the activities performed by them. Method: A survey of 105 professionals from 24 states collected data on organizational profiles, architect roles, compensation, and architectural tasks. Quantitative and qualitative analyses were conducted. Results: Architectural work is widespread but unevenly formalized. Only 41% of companies employ formal architects, and 68% delegate architectural duties to professionals without the title. Many respondents could not report how many architects their companies employ, indicating low internal visibility. Organizations whose core business is software development tend to have more architects, yet responsibilities are frequently shared across roles. Reported activities span design, technical leadership, mentoring, strategy alignment, and innovation. Formal architects typically receive higher salaries than informal ones. Conclusions: Software architecture in Brazil is practiced across diverse contexts but lacks consistent role definition. The findings highlight the need for clearer architectural roles, stronger organizational processes, improved educational preparation, and further research on the professionalization of architecture practice in Brazil.
Download

Paper Nr:	49
Title:	A Subway Map of Software Architecture: A Multivocal Review and Visual Framework
Authors:	Manoel Valerio da Silveira Neto, Andreia Malucelli and Sheila Reinehr
Abstract:	Understanding software architecture in agile environments remains challenging because architectural knowledge is dispersed across scientific, industrial, and normative sources. This paper presents the Subway Map of Software Architecture (SMoSA), a visual knowledge framework derived from a Multivocal Literature Review that integrates peer-reviewed studies, software engineering standards, and qualified gray literature. SMoSA organizes architectural practices, decisions, patterns, and quality attributes into ten thematic lines and interconnected stations, offering a structured visual representation of software architecture knowledge for agile contexts. The review followed a structured protocol for source selection, categorization, and evidence synthesis, and assigned confidence levels to support transparency about the nature of the supporting references. Finally, we report a preliminary expert-based evaluation to provide initial evidence of content adequacy and perceived utility.
Download

Paper Nr:	67
Title:	A Recommender System for Automated Security Threat Elicitation in Enterprise Information Systems
Authors:	Yuri Feitosa Negócio, Juliana Dantas R. V. de Medeiros and Nadja da Nobrega Rodrigues
Abstract:	Several threat modeling methods have been proposed in both industry and academia to enable early identification of security flaws in enterprise information systems. Nevertheless, their adoption by software development teams remains limited, mainly due to the required effort and the need for specialized expertise. In this context, this paper presents Threat Copilot, a recommender system that supports the automated elicitation of threats through the reuse of knowledge derived from previously developed organizational threat models. The tool was evaluated through offline and online validations, as well as a user-centered study based on the TAM-TTF acceptance model. The results demonstrate that the system is capable of recommending relevant threats, achieving precision, recall, and F1-measure values of 51%, 72%, and 56% in the offline evaluation, and 60%, 75%, and 67% in the online evaluation, respectively. In addition, the user-centered study indicated a reduction in the effort required for threat elicitation and increased consistency in the threat modeling process.
Download

Paper Nr:	71
Title:	Simulation-Based Performance Specification and Sizing for Enterprise Search Systems
Authors:	Eric Fernandes Monteiro and Afonso Sales
Abstract:	In the context of Enterprise Information Systems (EIS), search components are critical for knowledge management but often suffer from late-stage performance failures due to inadequate capacity planning. To address this, this study introduces a framework for Performance Requirements Engineering rooted in Discrete Event Simulation (DES). We propose a model-based approach that acts as a performance digital twin, allowing architects to simulate the behavior of search engine infrastructures under varying loads before deployment. Using a Python/SimPy-based artifact, we modeled the interplay between query complexity, index size, and hardware resources to validate architectural sizing decisions. The experiments quantify saturation thresholds (e.g., at 500 req/s) and memory bottlenecks, demonstrating how simulation can be used to elicit precise Non-Functional Requirements (NFRs) and define verifiable Service Level Objectives (SLOs). This approach shifts performance validation to the early design phase, enabling cost-effective scalability analysis and robust specification of enterprise search architectures.
Download

Paper Nr:	72
Title:	Preliminary Evidence on the Use of LLMs in Software Project Management Activities
Authors:	Julia Engels Adame, Felipe Sphair, Maria Lydia Fioravanti, Regina Fabia Lopes de Albuquerque and Leo Natan Paschoal
Abstract:	Software project management is undergoing a transformative period driven by the rise of Large Language Models (LLMs). Although the literature extensively explores the technical potential of these tools (e.g., automated code generation and requirements specification), there is a lack of empirical evidence regarding their practical application at the management layer. This paper presents a survey conducted with project managers working in the Brazilian software industry. The objective was to investigate the adoption landscape, operational benefits, technical barriers, and the need for human supervision in industrial practice. The results indicate that while LLMs act as productivity catalysts in planning and communication tasks, their application in complex business logic remains limited by hallucination phenomena and context window constraints (token limits). The research reveals that complete trust in automation is nonexistent, requiring “constant vigilance” from professionals. This work represents an initial effort to understand the redefinition of the project manager’s role, which is transitioning from an artifact drafter to a critical supervisor of LLM-assisted processes.
Download

Paper Nr:	80
Title:	Unveiling the Use of Large Language Models in Software Project Management: A Systematic Mapping Study
Authors:	Maria Lydia Fioravanti, Julia Engels Adame, Felipe Gabriel de Souza Sphair, Regina Fabia Lopes de Albuquerque and Leo Natan Paschoal
Abstract:	The use of Large Language Models (LLMs) has rapidly expanded across several software engineering activities, including software project management. Despite this growth, the literature still lacks consolidated evidence on how LLMs are applied in practice, which project management activities they support, and what benefits and limitations are associated with their adoption. This gap hinders the formulation of strategic guidelines for the safe and effective use of these technologies in organizational contexts. This paper presents a systematic mapping study aimed at synthesizing existing scientific evidence on the use of LLMs in software project management. The study analyzes how LLMs have been applied, identifies the models most frequently reported, and examines the project management activities they support, as well as the benefits and limitations described in the literature. The results indicate a predominance of GPT-based models, mainly applied to planning, monitoring and control activities, with reported benefits such as productivity gains, task automation, and reduced cognitive effort, alongside limitations related to incorrect responses, hallucinations, and context constraints.
Download

Paper Nr:	116
Title:	Large Language Models in Information Systems Engineering
Authors:	José Palazzo Moreira de Oliveira
Abstract:	The advent of Large Language Models (LLMs) represents a turning point in the foundations of Information Systems Engineering. Beyond their technical significance, LLMs challenge long-standing assumptions about how information, representation, and knowledge are structured. This article examines how LLMs reshape the relationships among language, meaning, and system design, suggesting that their emergence requires a re-examination of the core principles of contemporary information systems. We explore how the logic of generative models both extends and disrupts traditional approaches to in-formation and representation. The discussion highlights the importance of building LLM-based systems on transparent, ethically sound frameworks that respect human-centred knowledge processes. Ultimately, the paper argues that LLMs should be understood not merely as tools for automation but as active agents that transform the foundations of information systems engineering.

Paper Nr:	143
Title:	LLM-Driven Ontology Learning: From Term Extraction to Upper-Level Categorization for Building Information System Models
Authors:	Martin Ströher, Thaís Schäfer Luiz, Eduardo Roemers-Oliveira, Lucas Valadares Vieira, Fábio Herbert Jones, Luiz Fernando De Ros and Mara Abel
Abstract:	Knowledge-intensive Information Systems support high-value corporate tasks. However, building knowledge models requires manual selection and semantic clarification of terminology — a costly bottleneck in enterprise ontology engineering. This work explores automating these foundational steps to reduce manual effort during conceptual modeling. We evaluate the capacity of general-purpose Large Language Models (LLMs) to extract terminology and classify it into upper-level meta-types without specialized fine-tuning. We introduce an empirical framework comparing four term-extraction strategies: LLM + Corpus, LLM Generated, Named Entity Recognition (NER) + Corpus, and TF-IDF, applied to 40 Brazilian Pre-Salt scientific articles. To standardize classification, we leverage LLMs to generate Natural Language Definitions (NLDs) for the extracted terms, explicitly using these NLDs as semantic pivots to improve categorization into a multi-layered hierarchy integrating Basic Formal Ontology (BFO), GeoCore, and GeoReservoir ontologies. The resulting artifacts underwent blind validation by domain experts. Results demonstrate that non-specialized LLMs produce higher semantic quality in term extraction and taxonomic categorization compared to traditional statistical methods. The LLM + Corpus pipeline proved the most robust strategy. Ultimately, this work presents a validated paradigm for modernizing conceptual modeling and bootstrapping domain taxonomies through strategic enterprise-scale automation during conceptual modeling.
Download

Paper Nr:	150
Title:	Comparison between Object-Oriented Metrics for Software Maintenance
Authors:	José L. F. da Silva Junior, Raquel C. Martins and Michel S. Soares
Abstract:	Software Maintenance is considered a complicated activity in the software development life cycle, mentioned by many software developers as a burden to be performed to keep legacy software running. Metrics are well-known as useful for better understanding the internal and external software structure, as well as a first step towards determining the software quality, and to help software maintenance. Currently, there are a variety of different metrics, which may be confusing for the software developer. For instance, when to use a metric, the final purpose, and the interpretation in terms of quality for each metric. Therefore, in this paper, a comparison between well-known object-oriented software metrics is proposed. As a result, we propose a discussion on when to use each metric and how each metric is useful for activities of software maintenance. In addition, the study shows that combining structural and cognitive object-oriented metrics provides a more comprehensive view of software maintainability than relying on a single metric perspective.
Download

Paper Nr:	172
Title:	Model-Based Requirements Engineering for Data Science: Adapting the REMO Technique to Derive User Stories from BPMN Models
Authors:	Anna Beatriz Marques, Sérgio Barbosa Filho, Amanda Sousa, Angelina Sousa, Rhenara Oliveira, José Florêncio Neto, José Antônio Macêdo and Rossana M. C. Andrade
Abstract:	Data-intensive systems pose new challenges for Requirements Engineering (RE), as requirements often emerge from exploratory analyses, evolve, and depend on the availability and quality of data. This paper contributes to the growing field of data-driven and model-based RE by extending the REMO (Requirements Elicitation oriented by business process MOdeling) technique to derive agile requirements from BPMN models. The proposed adaptation establishes a semi-formal mapping between BPMN elements and user story components, bridging process-oriented modeling and agile specification practices. The approach was applied and evaluated in a real-world Research, Development, and Innovation (R&D&I) project that developed a Big Data platform for a public-sector organization that manages fund for educational development. In this context, the BPMN-based data pipeline acted as a structured elicitation mechanism connecting organizational managers and data specialists, supporting traceable specification of data product requirements. The study resulted in an evolved set of REMO heuristics tailored to user stories, 14 validated user stories, and an updated BPMN model that reflects improved process understanding. This experience report contributes to model-based Requirements Engineering for data-intensive Enterprise Information Systems by (i) extending the REMO technique with heuristics tailored for user story derivation, (ii) reporting its adoption in a real-world, interdisciplinary project, and (iii) discussing practical implications, limitations, and future research directions.
Download

Paper Nr:	182
Title:	Agile Project Management: A Comparative Analysis of Learning Strategies for User Story Decomposition and Task Dependency Identification
Authors:	Fadwa Rhorba, Driss Allaki and Mohamed Dahchour
Abstract:	In agile software development projects, user stories capture stakeholder needs. In practice, they are decomposed into technical tasks, while task dependencies are identified to support planning. The cognitive effort required to manually perform these activities is time-consuming and error-prone. As project complexity increases, maintaining accuracy becomes increasingly difficult, leading to unreliable planning, delays, and cost overruns. Prior studies overlook task-level dependencies and treat decomposition and dependency analysis separately. In this work, we empirically investigate whether user story decomposition and task dependency identification can reinforce one another by comparing single-task, multi-task, and transfer learning scenarios under a common pre-trained Generative Artificial Intelligence (GenAI) model to assess the impact of joint learning, and knowledge transfer on performance. The model follows an encoder–decoder Transformer architecture and is fine-tuned on a publicly available dataset of user stories organized into 22 projects and augmented with derived task descriptions and dependency labels. Using task-appropriate evaluation metrics, the results show that joint learning benefits decomposition but not dependency identification, while transfer learning preserves the former but degrades the latter. This trade-off motivates a decision-support system for agile project management that uses joint learning to improve decomposition and task-specific training for more reliable dependency identification.
Download

Paper Nr:	189
Title:	From Fiction to Requirements: Transforming Speculative Ideas into Game Requirements
Authors:	Cassio Ceolin Júnior, Bhruno Roan Leifheit, João Paulo Merlugo Ferreira, Paulo Silas Severo de Souza, Gilleanes T. A. Guedes, Ricardo Vilela, Pedro Henrique Valle and Williamson Silva
Abstract:	The development of digital games is a multifaceted process that goes beyond technical implementation, requiring the creation of interactive, narrative, and playful experiences. Requirements Engineering (RE) is fundamental in aligning the development team’s vision with users’ expectations. However, unlike traditional systems, digital games also require eliciting requirements that encompass subjective aspects, such as emotions, aesthetics, interaction, and immersion - elements that are difficult to capture by traditional RE methods. This article investigates how the Design Fiction technique can support requirements elicitation in the context of digital games, focusing on the technical quality of the suggestions generated for formulating user stories. To this end, we conducted an empirical study that applied Design Fiction sessions in the design phase of a Survival Horror game. Participants interacted with immersive speculative scenarios, and we analyzed their contributions. As a result, we obtained 68 speculative ideas, of which 59 were considered relevant and structured into 28 game requirements. Subsequently, we conducted an evaluation in which these requirements were decomposed into more granular units and refined to improve their technical precision, resulting in 46 game requirements in total.
Download

Paper Nr:	195
Title:	Self-Admitted Technical Debt across Multiple Software Versions: A Study on the COVID-19 Pandemic and Development Activity
Authors:	Giulia Campos de Oliveira, Jorge Marques Prates and Rogério Eduardo Garcia
Abstract:	In this paper, we investigate how developers’ activity and self-admitted technical debt (SATD) evolved before, during, and after the COVID-19 pandemic in an industrial software project. We analyze software versions spanning six years (2018–2023), grouped into three periods: pre-pandemic (2018–2019), pandemic (2020– 2021), and post-vaccine (2022–2023). Using a mining software repository approach. The dataset was organized to answer three research questions: (i) changes in developer activity; (ii) the evolution of SATD occurrence; and (iii) how SATD is distributed among developers. The results show that developer activity did not scale linearly with team size: during the pandemic, the number of developers increased while the volume of edited lines decreased. SATD also declined across the three periods and remained concentrated in a small subset of developers. These findings highlight the importance of monitoring both activity patterns and individual contribution profiles when interpreting SATD in contexts of organizational change.
Download

Paper Nr:	227
Title:	From Information to Application: Usable Technological Design (UTD) for Innovation Decision-Making
Authors:	Dorothea Schneider, Mauritz Mälzer and Steffen Ihlenfeldt
Abstract:	The practical adoption data-driven technologies in organizational contexts remains limited. Prior research emphasizes human-centered considerations as one reason. However, focusing on usability at the level of interfaces or explainability has proven inadequate to capture the broader socio-technical conditions under which technologies become applicable, adoptable, and valuable. Focusing on Small and Medium Enterprises (SMEs) in the engineering domain, this paper introduces a hierarchical construct that conceptualizes usability as a structured decision logic spanning suitability, compliance, and usefulness. The concept introduces two key contributions: (1) a hierarchy of necessary conditions that govern technology adoption, and (2) a temporal perspective on when different usability-related aspects can be meaningfully evaluated during technology development.
Download

Paper Nr:	266
Title:	MOTIVE Gateway: Architectural Enforcement of AI Governance for Enterprise LLM Integration
Authors:	Amadou Sienou
Abstract:	When enterprises deploy Large Language Models in production, a governance enforcement gap opens: regulations such as the EU AI Act and ISO/IEC 42001 specify what must be accountable, but existing infrastructure - authentication, logging, API gateways - cannot demonstrate intent, enforce constraints, or produce the structured evidence these regulations require (European Union, 2024; ISO/IEC, 2023). Governance remains a policy document; the technical layer does not know it exists. We argue that closing this gap requires treating AI governance as architectural infrastructure, not process overhead. The MOTIVE Gateway implements this argument as a Policy Decision / Enforcement Point (PDP/PEP) middleware layer that requires explicit business intent before any LLM call is admitted, validates requests against organisational policies, routes high-risk interactions to human reviewers, and logs cryptographically signed evidence to an append-only store. The design adapts the Values–Criteria–Indicators–Observables (VCIO) evidence model (Hallensleben et al., 2020) for runtime use, binding regulatory values to five control objectives and measurable metrics. We evaluate the architecture against a 200-task corpus spanning eight regulated domains. Results from the proof-of-concept - not from live enterprise deployments - show structural improvements in trace completeness, manifest completeness, and audit reconstruction. Seventy-two percent of rejected requests were successfully corrected on resubmission, suggesting the enforcement mechanism supports compliance learning rather than blocking productive work. Boundary conditions and organisational prerequisites are discussed, followed by a research agenda identifying the empirical work required to validate the architecture in live enterprise settings.
Download

Paper Nr:	285
Title:	Aligning Requirements Engineering and Design Teams: An Industrial Experience Report
Authors:	Evilasio Costa Junior, Fabiana Marinho, Artur Moraes, Erisson Nunes, Barbara Ferreira, Sarah Pimentel, Leticia Rodrigues, Lara Braga, Wendley Souza, Rossana Andrade and Miguel Franklin de Castro
Abstract:	This paper presents an empirical study of a distributed software project, focusing on Requirements Engineering (RE) challenges. The study highlights issues in requirements elicitation, analysis, specification, and validation under communication constraints. Misunderstandings and ambiguities in requirements were the primary challenges, resulting in rework. To address these challenges, the project combined RE practices with UI/UX activities, including user-centered design artifacts, collaborative workshops, and iterative stakeholder validation cycles. These strategies promoted shared understanding and better traceability from stakeholder needs to system requirements. Results show reduced ambiguity, higher-quality requirements, and more efficient validation. The paper offers lessons learned and methodological recommendations for Requirements Analysis, UI/UX design, and Management in Information Systems projects.
Download

Paper Nr:	299
Title:	Design of Automotive Embedded Systems Using SysML and UML: A Case Study on Window Control
Authors:	Diego J. Santos and Michel S. Soares
Abstract:	Automotive Embedded Systems have undergone significant technological advancements, increasing the com-plexity of system design. Managing systems integrating software with hardware presents a considerable challenge for systems engineers. The methodology employed involves detailed modelling of the core components within the automotive window control system utilising a Model-Based Systems Engineering (MBSE) approach. Consequently, this study demonstrates the application of MBSE by using the Systems Modeling Language (SysML) and Unified Modeling Language (UML) for the design of the models. This combination provides a structured methodology for modelling complex automotive systems, ensuring consistency throughout system development and highlighting the benefits of this approach. This analysis not only improves system quality but also facilitates communication between stakeholders, establishing it as a valuable method in the development of modern automotive systems and contributing to the understanding and application of SysML in conjunction with UML for Embedded Systems Engineering.
Download

Paper Nr:	313
Title:	Explainable AI (XAI) in Decision Support for Mitigating Carbon Emissions in Agriculture: A Systematic Review
Authors:	Pedro José C. de Almeida, Davi Moljo Domingues, Regina Braga, José Maria N. David and Bárbara de M. Quintela
Abstract:	Context: Climate change and global warming pose a challenge for researchers. COP 30 reinforces this issue by setting targets for mitigating greenhouse gas emissions. In digital and precision agriculture, authors face a dilemma: on one hand, they criticize the simplicity of the IPCC formulas, which do not always offer high precision for measuring emissions; on the other hand, they see obstacles in the adoption of direct emission measurement practices or in the use of biophysical modeling systems, where costs and scalability problems make widespread adoption unfeasible. Machine Learning technologies partially address this dilemma by balancing ease of use, scalability, precision, and cost; however, they lack transparency and auditability of the models. In this sense, Explainable Artificial Intelligence (XAI) techniques emerge to address this problem by interpreting and explaining the functioning of ML models, with a focus on the application domain. Objectives: This article presents a Systematic Literature Review (SLR) to build a body of knowledge on how XAI applied to Machine Learning models assists decision-making in agriculture. Method: A SLR following the PRISMA 2020 guidelines was used to select studies. We analyzed the studies to answer the following research question: "How have Explainable AI (XAI) techniques been applied in Machine Learning models linked to carbon emissions in the agricultural context to reveal new explainable strategies for producers and policymakers, which would be inaccessible with black-box models?" Results: The final sample consisted of 12 studies. The tool used as a standard in the studies was SHAP (SHapley Additive exPlanations), associated with different models, with emphasis on Ensemble learning models. Several ways in which the tools effectively generated explanations and greater transparency were categorized and presented. Methodological innovations regarding the intended use of SHAP Values were detected. Furthermore, the study identified strategic decisions in the agricultural sector that could be supported by XAI, including those made by rural producers and public policymakers. Finally, the study maps possible future research topics that still require validation, exploration, and a broader overview.
Download

Paper Nr:	13
Title:	Microfrontend: A Survey on Industry Practices
Authors:	Luiz Felipe Cirqueira dos Santos, Mariano Florencio Mendonça, Marcus Vinicius Santana Silva, Shexmo Richarlison Ribeiro dos Santos, Edmir Queiroz and Fabio Gomes Rocha
Abstract:	This study presents the results of a survey conducted with 16 professionals in the software development field, aiming to understand the practical adoption of micro frontend (MFE) architecture. The data reveal that MFEs have been utilized in companies across various sectors, particularly in complex projects and distributed teams, offering benefits such as increased team autonomy, modularization, faster delivery, and component reuse, with a particular emphasis on the use of React and tools like Azure DevOps and GitHub Actions. Despite these advantages, challenges such as governance, visual consistency, module communication, and security were found. Adopting MFEs demands maturity, standardization, and careful planning, and is not suited for small teams or simple projects. The full dataset is available at https://drive.google.com/drive/folders/1iM16rfeG5VRQ1OQ0SZfeE77WP4L-QR9q?usp=sharing.

Paper Nr:	70
Title:	Image Reshaping of Lexical and DNS Features for Malicious Domain Classification: A Stacking Ensemble Based on Visual, Textual and Tabular Models
Authors:	Luís Henrique Salomão Lobato, Adriano Mauro Cansian, Marcelo Zanchetta do Nascimento, Guilherme Freire Roberto, Sérgio Augusto Pelicano Junior and Leandro Alves Neves
Abstract:	To address the increasing sophistication of cyber threats, this study proposes a multimodal framework for malicious domain classification that integrates Machine Learning, NLP, and Computer Vision. We investigate image reshaping strategies Sequential Reshaping, Gramian Angular Fields (GAF), and Recurrence Plots (RPLOT) to transform lexical domain names and DNS features into visual representations. These are processed by ResNet-18, ConvNeXt-Nano, and ViT-B/16 architectures, and systematically benchmarked against NLP models (BERT, DistilBERT, DeBERTa) and tabular classifiers (Random Forest, XGBoost). Experimental results reveal that visual representations are highly competitive: ConvNeXt-Nano applied to GASF images achieved 85.50% accuracy, outperforming the strongest NLP baseline (DeBERTa, 84.13%). Building on these findings, was develop a stacking ensemble that integrates the best-performing visual, textual, and tabular models. This multimodal strategy delivers the most robust performance, reaching 89.10% accuracy and 95.50% AUC-ROC, confirming that the complementary integration of structural, lexical, and statistical features significantly enhances malicious domain detection.

Paper Nr:	85
Title:	Representing Dispositions with Vectors: A Case in Natural Gas Adsorption in Petroleum Production
Authors:	Cauã Roca Antunes, Mara Abel and João Cesar Netto
Abstract:	Representing dispositions is a core issue in modeling how entities participate in events and change throughout them. However, traditional approaches to representing dispositions bring significant limitations, particularly in engineering and industrial domains in which we must be able to associate the semantic and ontological models to existing mathematical frameworks. We present an alternative approach to representing dispositions, based on associating dispositions to potentials of change over qualities. Those potentials are represented by vectors that are called the impetus of dispositions. We apply this framework to model adsorption and desorption processes in natural gas dehydration, focusing on the changes over the water concentration of a portion of natural gas and a desiccant bed. We model the relevant dispositions, and their interactions are linked to an established mathematical model, the Langmuir adsorption equations. The approach demonstrates how ontological models can be tightly integrated with mathematical descriptions of physical processes, capturing key aspects such as reversibility, dynamic equilibrium, and the interplay between adsorption and desorption, enhancing semantic precision for scientific and engineering applications.
Download

Paper Nr:	95
Title:	XFrameTool: A Supporting Tool for Experimental Frameworks
Authors:	Leo Natan Paschoal, Tayana Conte and Simone Rocio Senger de do Souza
Abstract:	Variable selection is crucial in a controlled experiment, since variables must adequately represent the constructs of a domain. The lack of care in such a selection may threaten construct validity and risks arise when researchers must represent theoretical constructs using manipulable and measurable variables operationally. Those threats eventually impact the reliability of conclusions drawn from experiments and their comparisons. Towards dealing with variable selection, software engineering researchers have developed experimental frameworks that define manipulable and observable variables, establish precise definitions, indicate causal relationships, and demonstrate experimental designs incorporating those variables. However, they are typically presented as static documents (e.g., reports and scientific articles), requiring researchers to devote significant efforts to comprehend cause-effect relationships. A dynamic tool is necessary for supporting researchers in variable selection and mitigating threats to construct validity, facilitating experimental frameworks’ management, availability, and visualization. Additionally, no open-source software tool suitable for that purpose is currently available. This paper introduces XFrameTool, a novel web-based tool that fills such a gap by enabling the management, availability, and visualization of causal relationships within an experimental framework.
Download

Paper Nr:	141
Title:	Structuring a Development Process for AI/ML Projects: A Look into Industry Driven Issues
Authors:	Felipe Sonntag Manzoni, Leonardo Carneiro Marques, Camilla Rosas Gomes, Rayssa Campos dos Reis and Ana Carolina Oran Rocha
Abstract:	AI and ML projects differ in several aspects from traditional software development, so the traditional and agile development process cannot overcome the number of difficulties and aspects of these types of software development. In many cases, the AI and ML development cycles, even though representing just a fraction of the system, are made using non-standardized and inadequate development processes that cannot guarantee the developed product’s quality and effectiveness. This paper intends to elucidate and shed light on this issue, showcasing a novel development lifecycle process for AI/ML contexts that accounts for the quality activities and results to improve the project results. This research is developed in collaboration with the industry and is the first contribution on the developed process grounded on empirical information and validated on an empirical quality process. We present the first version of the development lifecycle process for AI/ML enabled systems and also present a initial evaluation of the developed artifact from the view of industry specialists, presenting the elicitation of further research and development for the next iteration.
Download

Paper Nr:	155
Title:	Collaborative Construction of Domain Ontologies Supported by Evidence-Grounded Large Language Models
Authors:	Silvia Lucia Borowicc, Wesley Lourenco Barbosa and Solange Nice Alves-Souza
Abstract:	Domain ontology construction remains a complex and resource-intensive task that must be both conceptually accurate and auditable over successive refinements. Although Large Language Models (LLMs) can accelerate content acquisition, their outputs may be unstable, insufficiently grounded, or inconsistent with modeling constraints, which motivates controlled use and explicit human oversight. This paper presents an end-to-end workflow for collaborative domain ontology construction that integrates structured terminological elicitation through the Language Extended Lexicon (LEL) to derive traceable seed-terms. The approach utilizes scoped LLM-assisted document analysis constrained by Retrieval-Augmented Generation (RAG) and few-shot prompting to anchor suggestions in authorized evidence. Human curation is performed within the OntoVis, a unified workspace developed by the authors as part of this work and presented as one of its contributions, which maintains a comprehensive, versioned change history to support review and auditing. Each iteration concludes with a validation suite acting as a compliance-ensuring gate to logical and structural integrity and functional adequacy. Demonstrated within the dengue surveillance domain, the workflow provides evidence that LEL-derived seeds and embedded validations enhance transparency and maintainability in human-in-the-loop ontology engineering.
Download

Paper Nr:	158
Title:	Inside Hackers' Minds: A Qualitative Study of Offensive Security Tool Developers
Authors:	Nestori Syynimaa
Abstract:	Offensive Security Tools (OSTs) are widely used in the cybersecurity industry by both defenders and attackers. Due to potential negative consequences, the ethicality of publishing OSTs is widely debated. OSTs are often published by so-called White Hat hackers, who generally aim to do good. This raises the question: why publish something that can be used to cause harm? In this qualitative research, we seek to understand the background, motivation, and ethical views of OST developers. For this purpose, ten well-known OST developers were interviewed. The findings show that OST developers do carefully consider ethical issues when developing and publishing their tools. They are motivated by non-financial factors, such as making the world a safer place. Finally, OSTs seem to have a strong positive professional consequences to their developers.
Download

Paper Nr:	159
Title:	A Framework for Sustainability and Code Smells Identification
Authors:	Jone Correia, Luís Rivero and Davi Viana
Abstract:	Software sustainability and code quality are increasingly relevant in Information Systems (IS) development, yet existing approaches do not systematically connect sustainability analysis with code smell identification. This paper presents the CSCS framework, a Design Science artifact that extends SusAF through a structured Notebook for mapping code smells across impact order and five sustainability dimensions: social, environmental, economic, individual, and technical. A pilot evaluation with 12 ICT professionals provided initial evidence of the framework’s usefulness and practical applicability. By linking sustainability awareness to code smell identification, the framework can support maintainability, reduce rework, extend system lifespan, and foster organizational learning around sustainable information systems.
Download

Paper Nr:	200
Title:	Milestone: A Neural Network Based Framework to Continuous Monitoring and Evaluation of Performance in Self-Adaptive Systems
Authors:	Belmondo R. A. Junior, Rossana M. C. Andrade, Valeria L. L. Dantas, Rainara M. Carvalho, Tales P. Nogueira and Marcio E. F. Maia
Abstract:	Self-adaptive systems (SAS) continuously monitor, analyze, plan, and execute adaptations to sustain quality under changing conditions. A persistent challenge is turning non-functional requirements (NFR), especially performance, into runtime evidence that can be observed and evaluated to support adaptation decisions. To mitigate that challenge, we present Milestone, a framework that provides continuous monitoring and evaluation of performance for SAS by replacing brittle rule-based assessment with a neural-network–supported component. Milestone collects runtime data through lightweight software sensors, classifies performance states, and supports adaptation planning and knowledge reuse. We evaluate Milestone using the Google 2019 Cluster Sample, reporting classification effectiveness via accuracy, ROC-AUC, and loss, and assessing planning through latency and success rate. In our experiments, the model achieved ROC-AUC = 0.86 and loss = 0.43, while planning latency remained below 1 second with a 66.67% success rate. Finally, a proof of concept integrates Milestone with SA-BSN (Self-Adaptive Body Sensor Network), showing how detected degradation can trigger adaptation in a cyber-physical exemplar.
Download

Paper Nr:	265
Title:	A Governance-Oriented Structured Diagnostic Approach for Budgetary Process Improvement: Evidence from a Brazilian Electoral Court
Authors:	Marilia Cristina Sassim Jesus, Edna Dias Canedo, Daniel Alves da Silva, Guilherme Fay Vergara, Paulo Tadeu Moreira Saldanha and Fábio Lúcio Lopes de Mendonça
Abstract:	Context: Public administration institutions face increasing pressure to improve efficiency, transparency, and accountability in budgetary governance. Although Business Process Management (BPM) has been widely adopted in public organizations, structured diagnostic approaches specifically tailored to budgetary processes in high-accountability contexts remain underexplored. Goal: This study aims to investigate how a structured diagnostic approach can systematically assess and improve budgetary processes in high-accountability public institutions. Method: A single embedded case study was conducted in a Brazilian Electoral Court. The research integrates BPMN-based AS-IS modeling, governance-oriented diagnostic assessment, and impact– effort prioritization within a structured four-phase framework. Data were collected through workshops, documentation analysis, and process walkthroughs, ensuring triangulation and analytical rigor. Results: The findings reveal that coordination density, validation intensity, documentation fragmentation, and traceability gaps constitute structural drivers of inefficiency in the budget proposal phase. These fragilities propagate across execution and monitoring stages. The impact-effort analysis shows that 76.47% of identified improvement opportunities correspond to medium-to-high impact initiatives with low-to-medium implementation effort. Conclusions: The results demonstrate that structured diagnostics function as governance instruments, enabling evidence-based prioritization and procedural refinement. The proposed framework offers a replicable approach for strengthening efficiency and accountability in public budgetary processes.
Download

Paper Nr:	284
Title:	Requirements Engineering in E-Government Portal Development: A Systematic Mapping Study
Authors:	Laryssa Ribeiro Oliveira, Alinne Souza, Williamson Silva and Thiago D. Cordeiro
Abstract:	E-Government refers to the use of information and communication technologies to improve the delivery of public information and services. Although digital government initiatives can increase efficiency, transparency, and citizen engagement, designing high-quality e-government portals that meet citizens’ needs while complying with quality and regulatory requirements remains a major challenge. In this context, Requirements Engineering (RE) plays a central role, particularly in supporting user-centered design, requirements specification, and adherence to quality standards. This paper presents a Systematic Mapping Study (SMS) on RE in e-government portal development to characterize the state of the art. The results indicate a predominance of evaluation-oriented contributions, with a strong emphasis on user-centered quality concerns (especially accessibility and usability). The literature also reports recurring organizational, technical, and external challenges, while showing limited use of experiments and relatively few transferable prescriptive artifacts (e.g., guidelines and techniques). Overall, this study provides an empirically grounded overview of how RE has been addressed in e-government portals and highlights opportunities for stronger cumulative evidence and more reusable RE guidance.

Area 4 - Software Agents and Internet Computing

Full Papers

Paper Nr:	89
Title:	AI Applications on RISC-V and ARM Architectures: A Focus on Fog Computing, Edge Computing, and IoT
Authors:	Andre Reges Souza Meira, Edward David Moreno and Calebe Conceição
Abstract:	This paper presents a comprehensive survey of RISC-V and ARM architecture applications for AI workloads across fog computing, edge computing, and IoT domains. We analyze recent developments (2020-2025) in instruction set extensions, neural processing units, and software frameworks that enable efficient machine learning inference on resource-constrained devices. Our comparative analysis reveals that RISC-V offers superior energy efficiency (up to 4× lower power consumption) and architectural flexibility through custom ISA extensions, while ARM maintains advantages in ecosystem maturity, raw inference performance (up to 15× faster for large networks), and commercial deployment readiness. The survey covers AI workloads spanning computer vision, natural language processing, anomaly detection, and sensor fusion, with a particular focus on TinyML implementations that achieve sub-milliwatt operation. We identify emerging trends, including inmemory computing, neuromorphic architectures, and heterogeneous computing approaches that combine CPU cores with domain-specific accelerators. Our findings indicate that architecture selection critically depends on application requirements, with RISC-V preferred for power-constrained custom acceleration and ARM for performance-critical deployments requiring mature toolchain support.
Download

Short Papers

Paper Nr:	90
Title:	Privacy and Artificial Intelligence in Brazil: An Analysis of Bill 2,338/2023 and Its Complementarity with the LGPD
Authors:	Luiz Henrique Ferreira E. Pereira, Stefano Luppi Spósito, Fábio Lúcio Lopes de Mendonça and Edna Dias Canedo
Abstract:	Brazil’s regulatory debate on Artificial Intelligence (AI) has culminated in Bill No. 2,338/2023, which establishes a national legal framework for AI. This paper examines how the Senate approved version structures privacy and personal data protection governance and how it complements Brazil’s General Data Protection Law (LGPD). We conduct a qualitative documentary analysis of the consolidated bill text and key amendments, triangulated with international reference frameworks, including the European Union AI Act, OECD principles on AI and data governance, and selected regulatory approaches from China and the United States. The analysis identifies a three-layer normative architecture: (i) a principle based layer grounded in fundamental rights; (ii) a procedural layer establishing rights to information, explanation, contestation, and meaningful human review; and (iii) a governance layer institutionalizing the Algorithmic Impact Assessment (AIA) as a preventive risk-management instrument and creating the National AI Regulation System (SIA), coordinated by the National Data Protection Authority (ANPD). Results indicate that the bill adopts a risk-based regulatory approach that is largely complementary to the LGPD, aiming to reduce normative overlap while consolidating data protection as a central pillar of responsible AI governance. We conclude that Bill 2,338/2023 represents a substantial step toward harmonizing innovation with safeguards for individual rights in Brazil’s digital ecosystem.
Download

Paper Nr:	164
Title:	Instruments and Artifacts for Evaluating Educational Chatbots
Authors:	Mariana Almeida Faustino, Leo Natan Paschoal and Pedro Henrique Valle
Abstract:	The use of educational chatbots has been growing significantly as support for teaching and learning processes. Despite this expansion, the field lacks methodological consensus on how to evaluate these systems in a systematic, rigorous, and comparable manner. Therefore, the objective of this work was to identify and characterize the instruments used in the literature for the evaluation of educational chatbots, mapping them against the ISO/IEC 25010 software quality standard. To this end, we conducted a Systematic Mapping Study (SMS) in four major databases (ACM Digital Library, IEEE Xplore, Scopus, and Web of Science), without date restrictions, resulting in a final sample of 32 primary studies. Results identified 41 evaluation instruments. The analysis reveals a critical asymmetry: evaluation is heavily skewed towards Quality in Use (e.g., Satisfaction, Acceptance), using domain-independent scales like TAM and SUS, while Product Quality attributes (e.g., Reliability, Security) are largely neglected. A methodological dichotomy was observed: text-based chatbots are evaluated as utilitarian tools, whereas multimodal agents focus on social presence. We conclude that the field prioritizes the student’s external perception over the software’s internal robustness and lacks methodological consistency, as studies either apply standardized instruments directly or rely on unvalidated adaptations, highlighting an urgent need for specialized frameworks that reduce dependence on ad hoc modifications.
Download

Paper Nr:	219
Title:	An Empirical Comparison of CNN and ViT Features for Visually-Aware Fashion Recommendation
Authors:	Lucas Silva Couto, Rodrigo Calvo, Rodrigo Clemente Thom de Souza and Marcos Aurelio Domingues
Abstract:	With the growing volume of visual information in fashion e-commerce, it is challenging to effectively recommend new clothing items to users. Visually-aware recommender systems can address this issue by incorporat-ing image-based features, typically extracted using deep learning models such as Convolutional Neural Networks (CNNs) or Vision Transformers (ViTs). However, there is still limited empirical evidence on whether hybrid combinations of CNN and ViT features can consistently improve visually-aware recommender systems. In this work, we investigate the impact of different visual feature extractors on fashion recommendation. Using a large-scale subset of the Amazon Reviews 2023 dataset, we compared ResNet50, ViT-B/16, and a hybrid CNN–ViT representation as visual backbones for three representative visually-aware models (i.e., Visual Bayesian Personalized Ranking (VBPR), Visual Neural Personalized Ranking (VNPR), and DeepStyle) under a top-10 recommendation setup. Overall, ViT-based features achieved a performance close to the ResNet50 baseline across all evaluated metrics. Regarding our proposal, the hybrid CNN–ViT embedding yielded modest but consistent gains in VNPR, while it does not provide measurable benefits for VBPR and DeepStyle, where ResNet50 remains the strongest option. These findings suggest that, in the evaluated scenario, naive Principal Component Analysis (PCA)-based fusion of CNN and ViT features increases representation complexity without delivering reliable ranking improvements, and that well-established single-backbone architectures remain a competitive and practical choice for visually-aware fashion recommendation.
Download

Paper Nr:	310
Title:	Systematic Mapping Study on Requirements Negotiation Techniques: An Update and Extension
Authors:	Lincoln Balbiano Pereira, Gilleanes Thorwald Araujo Guedes and Williamson Silva
Abstract:	Requirements negotiation plays a central role in Requirements Engineering, as software projects frequently involve multiple stakeholders with conflicting goals, priorities, and constraints. Managing these conflicts effectively is essential to ensure alignment between business objectives, technical feasibility, and stakeholder expectations. Considering this context, Tito et al. (2017) conducted the first systematic mapping study dedicated specifically to requirements negotiation techniques, identifying ten approaches predominantly grounded in collaborative negotiation paradigms. However, since the publication of that study, the landscape of software engineering has evolved considerably, with the emergence of new decision-support models, computational negotiation mechanisms, and Artificial Intelligence techniques. Our study replicates and extends the mapping conducted by Tito et al. (2017) by expanding the temporal scope to include studies published between 2017 and 2025. The results identify eighteen techniques reported in recent literature and highlight the emergence of hybrid, multicriteria, and AI-supported approaches, including Large Language Models (LLMs). The findings demonstrate a transition from purely collaborative negotiation models toward computational and structured decision-support approaches, indicating a significant evolution in how requirements negotiation is supported in contemporary software engineering environments.
Download

Paper Nr:	53
Title:	Data Security Challenges in the Quantum Computing Transition: Privacy Implications for the Brazilian Public Sector
Authors:	Loyane Mota Fernandes, Stefano Luppi Spósito, Fábio Lúcio Lopes de Mendonça and Edna Dias Canedo
Abstract:	Context: Digital transformation in government, driven by technologies such as the Internet of Things (IoT), artificial intelligence, and cloud computing, has significantly increased the exposure of sensitive data and challenged traditional security models. With the advent of quantum computing, classical cryptographic algorithms are expected to become vulnerable, requiring new strategies for data protection and governance. Goal: This study investigates how public policies and regulatory frameworks can adapt to post-quantum threats, with a focus on privacy, security, and accountability in the Brazilian federal government. Method: A qualitative and exploratory approach was adopted, comprising three stages: (i) a systematic literature review (2020–2025) conducted in accordance with the PRISMA 2020 guidelines, (ii) a comparative analysis of international regulations LGPD, GDPR, and HIPAA and (iii) a case study of the Brazilian federal government based on the results of the 2023 and 2025 self-assessment reports conducted by the Ministry of Management and Innovation in Public Services (MGI). Results: The review identified 23 primary studies, grouped into three main dimensions: (i) technological infrastructure, focused on post-quantum protocols and continuous authentication; (ii) governance and regulation, emphasizing the integration of innovation and compliance; and (iii) advanced protection methods, such as lattice-based cryptography, quantum key distribution (QKD), and blockchain. GDPR remains the most prescriptive framework, while LGPD adopts an intermediate approach and HIPAA follows a sector-specific model. In the federal government context, the findings indicate an early and uneven stage of organizational maturity, with significant gaps in cryptographic inventory, key management, and security event correlation. Conclusion: The study demonstrates that building a quantum-resilient governance model in the public sector requires the integration of validated technical mechanisms, adaptive regulation, and an institutional culture of accountability. A structured transition roadmap aligned with NIST CSF 2.0 and LGPD is proposed, encompassing cryptographic inventory, hybrid algorithm adoption, crypto-agility, and continuous capacity building. The evidence suggests that quantum-regulatory resilience in the public sector is not a one-time effort, but an evolving process that integrates science, regulation, and institutional practice to ensure security and trust in the post-quantum era.

Area 5 - Human-Computer Interaction

Full Papers

Paper Nr:	43
Title:	A Review on the Integration of Usability and Scrum in Software Development
Authors:	Vanessa Matias Leite and Marcelo Morandini
Abstract:	The increasing adoption of agile methodologies has driven the pursuit of more efficient software development processes. Nevertheless, challenges remain regarding the integration of usability practices, particularly within the Scrum framework, which traditionally emphasizes the rapid delivery of functionalities. The absence of clear guidelines for incorporating user experience (UX) aspects into agile practices reveals a gap between user needs and the outcomes delivered by development teams. This systematic literature review aims to examine how usability practices have been adopted, adapted, and evaluated within the context of Scrum, providing insights for researchers and practitioners seeking to align user experience with agile principles. The study followed formal guidelines for systematic literature reviews, applying a search protocol to the Scopus and Web of Science databases. The search terms included concepts related to Scrum, agile development, human–computer interaction (HCI), and UX. After applying inclusion and exclusion criteria, 24 primary studies were selected for qualitative analysis. The analyzed studies reveal multiple strategies for integrating usability into Scrum. The most frequently employed artifacts include prototypes, user stories, personas, and mockups. The main evaluation practices identified were user testing and Nielsen’s heuristics. Recurring challenges include weak collaboration between UX professionals and developers, the absence of usability requirements in sprints, and resistance from design teams to adopting agile approaches. The integration of usability into Scrum varies considerably across the reviewed studies, both in terms of the artifacts employed and the evaluation strategies adopted. These variations reflect the lack of consolidated guidelines and the uneven adoption of user-centered practices in agile environments, underscoring the need for more clearly defined and integrated models.
Download

Paper Nr:	66
Title:	Architectures and Frameworks for Collaborative XR Spaces: A Systematic Literature Review
Authors:	Gustavo Machado de Freitas, Cesar Tadeu Pozzer and Lisandra Manzoni Fontoura
Abstract:	The transition from physical horizontal surfaces, such as tabletops, to immersive interactive spaces in Extended Reality (XR) expands opportunities for collaborative interaction, shared perception, and decision support. This shift also introduces challenges related to Human–Computer Interaction (HCI), software engineering, and user coordination in collaborative systems that rely on multiple Head-Mounted Displays (HMDs). This study investigates the architectures, frameworks, and implementation strategies adopted in the development of collaborative interactive surfaces and spaces in XR, focusing on how technical design decisions shape interaction, co-presence, and synchronous collaboration among users. We conducted a Systematic Literature Review following Kitchenham’s protocol and the PRISMA guidelines. We searched the IEEE Xplore Digital Library, ACM Digital Library, and Scopus databases and considered studies published between 2015 and 2025. After screening and selection, we analyzed 39 primary studies addressing network architectures, synchronization techniques, concurrent interaction mechanisms, user representation strategies, and reported technical challenges. The analysis reveals a predominance of client–server architectures based on the Unity engine and network middleware, particularly Photon. Many systems address spatial and state synchronization separately, affecting the consistency of shared experiences and collaborative interaction. These findings highlight the need for more robust architectures, wider adoption of open standards, and integrated evaluation.
Download

Paper Nr:	86
Title:	Last Flight of Maverick: An Exploratory Qualitative Interview Study of GitHub Copilot Use in Software Development Team
Authors:	Sérgio Cavalcante, Edierley Messias, Erick Ribeiro and Ana Carolina Oran
Abstract:	Generative AI (GenAI) code assistants such as GitHub Copilot are increasingly adopted in professional software development. While prior work reports productivity and quality effects using controlled experiments or large-scale quantitative evidence, fewer studies provide a dense account of how practitioners experience these tools in real organizational settings, including trade-offs in trust calibration, learning, workflow friction, and dependency. This paper reports an exploratory qualitative interview study with five practitioners (junior, mid-level, senior developers, and a Qquality Analyst(QA)) from a single industrial team using GitHub Copilot in day-to-day work. Findings are organized into six analytic dimensions (productivity, perceived code quality, learning, challenges/limitations, dependency, and satisfaction) and interpreted through a Developer Experience (DevEx) lens. Participants reported substantial perceived productivity gains in repetitive tasks and test-related work, frequent use of Copilot as a drafting and review partner, and meaningful learning support for unfamiliar technologies. At the same time, they highlighted the need for continuous validation, context limitations, IDE frictions, enterprise account constraints, and risks of over-reliance. We synthesize evidence-informed adoption recommendations and discuss threats to validity and implications for research and practice.
Download

Paper Nr:	184
Title:	Chatbot Design: A Systematic Literature Review
Authors:	Antônio Silva, Andrey Pimentel, Tatiana Riechi, Thaís Fogaça and Marta Ferreira
Abstract:	This study presents a Systematic Literature Review (SLR) on chatbot design, aiming to characterize the state of the art, identify recurring challenges, and uncover research gaps and opportunities. The review follows the Goal–Question–Metric (GQM) approach and Kitchenham’s guidelines to ensure methodological rigor. Searches were conducted across six scientific databases, resulting in 1,567 records, of which 19 peer-reviewed studies were included after a two-stage screening process. The findings reveal a predominance of rule-based and flow-driven approaches, widespread use of mobile platforms, heterogeneous adoption of development frameworks, and limited reporting of security and privacy practices. Although social, cultural, and psychological factors are frequently considered, they are typically incorporated without formal structuring or explicit modeling mechanisms. No studies were found that employ formal ontologies or the biopsychosocial model as organizing dimensions. Overall, the review highlights gaps in semantic modeling, data governance, and explainability, pointing to the need for more systematic and transparent design approaches, especially in applications targeting vulnerable populations.
Download

Paper Nr:	202
Title:	SmartPATHY: A Prompt-Based Methodology for Empathy-Aligned Persona Generation with Large Language Models
Authors:	Ganiyat Saleeman, Ana Carolina Oran and Bruno Gadelha
Abstract:	Personas are widely used in software engineering and human-centered design to support understanding of user needs and guide design decisions. However, creating high-quality personas is often time-consuming and cognitively demanding. Recent advances in Large Language Models (LLMs) have opened new possibilities for supporting persona creation through prompt-based interaction, but questions remain regarding usability, reliability, representativeness, and alignment with structured persona frameworks. This paper presents and empirically evaluates SmartPATHY, a prompt-based methodology for generating high-quality personas aligned with the PATHY framework using Large Language Models (LLMs). Although high-quality personas are traditionally grounded in empirical user data, this study does not position LLMs as a substitute for user research. Instead, representativeness is supported through a structured, PATHY-aligned prompt workflow that explicitly incorporates predefined user needs, public audience characteristics, and empathic dimensions as design constraints.
Download

Paper Nr:	268
Title:	Integrating Gamification and Usability: A Framework for Usability Recommendations for Gamified Systems
Authors:	Kristina Magylaitė and Rimantas Butleris
Abstract:	Gamified systems incorporate gamification elements to influence user behaviour and engagement. Although gamification frameworks and usability standards exist, their integration remains insufficiently formalised. Gamification elements are often classified inconsistently across abstraction levels, while usability guidelines are typically applied at the system level without linkage to specific gamification elements. This paper proposes a framework for constructing and applying usability recommendations for gamified systems. The approach integrates a taxonomy of gamification elements organised according to the MDA model, a recommendation metamodel, and a tool for determining usability recommendations for gamified systems. The framework generates usability recommendation sets based on parameters such as age group, application domain, gamification goals and usability goals. Its application to three systems from different domains demonstrates that the framework produces different recommendation sets depending on user and system parameters. The proposed framework provides a formal approach for linking gamification elements with usability recommendations.
Download

Paper Nr:	271
Title:	Nudging Away from Online Extremism: A Review of Digital Nudges as Tools for Polarization De-Escalation
Authors:	Wladimir A. P. Neves, Angelica Dias, António Correia and Daniel Schneider
Abstract:	Political polarization and misinformation on social media platforms are phenomena deeply driven by the choice architectures and algorithmic logics of digital platforms, which systemically foster echo chambers and filter bubbles. To investigate how behavioral economics-based interventions can mitigate this informational crisis, this study presents a Systematic Literature Review encompassing 12 peer-reviewed articles published between 2019 and 2025. The research proposes a tripartite analytical framework that categorizes Digital Nudging at the Interface (Salience), Process (Friction), and Backend (Algorithmic Structure) levels. The results consolidate the Interface vs. Structure Paradox: while visual interface nudges effectively curb impulsive sharing by engaging System 2 cognition, their depolarizing potential is limited by the "Boomerang Effect." In contrast, structural backend interventions—with emphasis on Random Dynamical Nudges (RDN)—exhibit significant technical efficacy in breaking homophily and deconstructing echo chambers at the root, yet they operate opaquely as an "invisible hand," bordering on ethical dilemmas associated with Dark Patterns. It is concluded that the mitigation of extreme polarization transcends unidimensional solutions, requiring personalized interventions that orchestrate a non-negotiable symbiosis between the efficacy of algorithmic steering (Nudge) and the metacognitive design of empowerment and transparency (Boost), thus preserving the autonomy of the digital citizen.
Download

Paper Nr:	276
Title:	Improving Interaction with Physical Objects in AR Assembly Training through Spatial Synchronization
Authors:	Gabriel Di Domenico, Gustavo Machado de Freitas, Lisandra Manzoni Fontoura and Cesar Tadeu Pozzer
Abstract:	In the context of Industry 4.0, training novice operators for complex manual assembly tasks remains a significant challenge for enterprise systems. A major difficulty in Augmented Reality (AR) training environments is the interaction gap between physical objects and their virtual representations, which often requires manual alignment and increases cognitive workload. This paper proposes an interaction technique based on real-time spatial synchronization between physical objects and their Digital Twin counterparts in AR-assisted assembly training. By leveraging vision-based pose estimation, the proposed approach automatically maintains spatial congruence between the manipulated physical object and its virtual representation, allowing users to interact naturally with physical components while receiving digital guidance. An empirical study (N = 19) was conducted comparing three instructional conditions: synchronized AR, manual AR, and traditional 2D instructions. Results show that synchronized AR significantly improves task performance, achieving a 34% reduction in task completion time compared to manual AR and a substantially lower perceived workload as measured by NASA-TLX. The findings indicate that spatial synchronization reduces the split-attention effect and interaction overhead by eliminating the need for manual viewpoint adjustments. These results suggest that spatial synchronization between physical objects and Digital Twins represents an effective interaction mechanism for AR-based enterprise training systems, improving usability and supporting more intuitive interaction with physical components.
Download

Short Papers

Paper Nr:	15
Title:	A Set of Good Practices for Accessible Product Management
Authors:	Letícia Gomes Macedo and Marcelo Medeiros Eler
Abstract:	Digital products increasingly mediate access to essential services, yet accessibility remains inconsistently embedded in product work. This paper proposes and validates a practitioner-oriented set of good practices to help product managers and product leaders integrate accessibility across the product lifecycle. We conducted a systematic review in Scopus, extracted software product management activities and practice statements, adapted them to an accessibility perspective, and synthesized them through constructivist grounded-theory coding. The resulting candidate practices were validated in a two-stage questionnaire with product practitioners using five criteria: Understanding, Attribution-to-Product, Importance, Feasibility, and Application. The final result is a validated set of 13 practices, including 9 for product managers and 4 for product leadership. Findings show high perceived importance and clarity, but lower application, suggesting that the main barrier is organizational support rather than lack of value recognition.
Download

Paper Nr:	42
Title:	Temporal SWIM: A Technique for VR Temporal Navigation Avoiding Loss of Context
Authors:	Gustavo Machado de Freitas, Natan Luiz Paetzhold Berwaldt, Gabriel Di Domenico, Lisandra Manzoni Fontoura and Cesar Tadeu Pozzer
Abstract:	In recent years, Virtual Reality has emerged as a powerful tool for analyzing complex spatiotemporal data in domains such as meteorological analysis, military training simulations, and immersive sports reviews. While spatial recordings allow users to navigate through both space and time to understand these dynamics, temporal transitions often disorient users due to the difficulty of maintaining spatial context. This issue compromises data comprehension and decision-making efficiency, especially in scenarios requiring continuous analysis. To address this, this work proposes Temporal SWIM, an innovative interaction technique that integrates the Scalable World-In-Miniature (SWIM) approach with an interactive timeline. The solution allows users to select regions of interest, visualize their changes over time in a miniature view, and navigate between temporal states without losing environmental context. The technique was validated through controlled experiments involving continuous spatial recordings. Results showed that Temporal SWIM significantly reduced disorientation and errors in analytical tasks, providing better accuracy and understanding of environmental changes compared to a traditional 2D timeline, thereby improving overall task efficiency.
Download

Paper Nr:	82
Title:	When Awareness Is Not Enough: An Experimental Study on Human Susceptibility to Phishing
Authors:	Gabriel Borges da Conceição and Edna Dias Canedo
Abstract:	Phishing remains one of the prevalent cybersecurity threats, primarily exploiting human factors rather than technical vulnerabilities. Despite advances in automated detection mechanisms, phishing attacks continue to succeed by manipulating users’ cognitive and behavioral traits. This paper investigates phishing susceptibility from a socio-technical perspective, focusing on the gap between users’ awareness of phishing risks and their actual behavior at the moment of decision-making. A controlled social experiment was conducted with 48 undergraduate students enrolled in computer-related programs, who possessed prior knowledge of cybersecurity concepts, simulating a phishing scenario. The results reveal that even technically trained individuals remain vulnerable, with 43.75% of participants clicking on a fraudulent link. These findings reinforce evidence that situational factors, such as haste and misplaced trust, can override technical knowledge and prior awareness. Based on the observed behaviors, this study derives design implications for lightweight, user-centered interventions that support human judgment in real time. As a proof of concept, a privacy-preserving browser extension illustrates how visual cues and contextual alerts can be integrated into webmail environments to mitigate phishing susceptibility without relying on centralized data processing. By framing phishing as socio-technical challenge, this work contributes empirical evidence on vulnerability and highlights the role of usable security mechanisms embedded in Information Systems.

Paper Nr:	88
Title:	The Oscar Goes to: Controversy? Analyzing Toxicity and Emotional Intensity During the Oscars on YouTube Live Comments
Authors:	Gabriela B. Kurtz, Erick B. Machado, Vinícius Ross W. dos Santos, Gabriel C. Bremm, Gabriel Z. Souza, Isabel H. Manssour, Roberto Tietzmann and Milene S. Silveira
Abstract:	Considering the potential of digital platforms to enable public discussions on any subject, these can quickly turn into toxic dialogues. To help understand this scenario, we developed and evaluated an approach for detecting and analyzing toxic discourse in Portuguese comments on YouTube live streams. The approach enables the integration of multilingual, domain-adapted models to enhance the reliability of toxicity classification in social media contexts. To test its applicability, we used the tool in the 2025 Academy Awards ceremony as broadcast on YouTube. In the 2025 edition, the Brazilian film ”Ainda Estou Aqui” (I’m Still Here) was nominated, and Brazilian fan bases were intensely mobilized, which in turn fueled the emergence of online toxic manifestations. As contributions, we present a comparative evaluation of toxicity detection models to increase confidence in identifying those comment types, a new fine-tuned toxicity detection model for Portuguese YouTube comments, and an interactive web-based tool to support the analysis of toxic discourse, which was used with the aforementioned case study.
Download

Paper Nr:	118
Title:	A Set of Evidence-Based Design Principles for AI-Supported Decision Making in Scrum Retrospectives
Authors:	Felipe Pritsch Fahrion, Afonso Sales, Rafael Chanin, Nicolas Nascimento and Leonardo Brenner
Abstract:	Scrum retrospectives are essential for continuous improvement yet frequently suffer from superficial reflection, weak follow-up, and psychological safety challenges. While Artificial Intelligence (AI) has been increasingly explored in agile contexts, its application to retrospectives remains underexplored, particularly from a human-centered and sociotechnical perspective. This paper argues that before evaluating the effectiveness of AI-based tools for retrospectives, it is necessary to clarify how such systems should be designed and adopted responsibly. To this end, we present a design-oriented investigation grounded in the triangulation of scientific literature, grey literature, and qualitative insights from industry professionals. Rather than proposing a concrete system, the study synthesizes this multivocal evidence into a set of eleven Design Principles that articulate constraints and orientations for AI-based decision support in Scrum retrospectives. The proposed principles frame retrospectives as a closed-loop decision process and position AI as an assistive, subordinate actor that supports information synthesis, pattern identification, and longitudinal learning, while preserving human facilitation, trust, and psychological safety. By contributing empirically grounded design knowledge, this paper aims to support future research and practice on human-centered AI for collaborative decision-making, providing a foundation for the development and empirical evaluation of AI-supported retrospective tools.
Download

Paper Nr:	130
Title:	Digital Inclusion of Older Adults: Challenges and Solutions for Brazilian Government Platforms
Authors:	Andrea Judice, Marcelo Judice, Glauco Pedrosa and Rejane Figueiredo
Abstract:	Population aging is a global phenomenon that also affects Brazil. With increasing life expectancy, the elderly population is growing, bringing new demands, including technological ones. The digital inclusion of the elderly is a relevant topic in the areas of technology, interaction design, and public policies. In this context, this study analyzed the challenges faced by elderly people (70 to 90 years old) in using Brazilian public platforms, such as Meu INSS and GOV.BR, focusing on usability, accessibility, and inclusion. The main barriers identified were difficulty with passwords, confusing navigation, and fear of fraud. Despite this, many elderly people showed interest in learning how to use the platforms, suggesting improvements such as simplified login, clearer language, and training. The results obtained reinforce the need to adapt these platforms to promote a more accessible and safe experience, giving the public sector the opportunity to improve digital inclusion and increase the autonomy of the elderly.

Paper Nr:	133
Title:	An Analysis of User Experience in Customer Service Chatbot Interactions Using the UEQ
Authors:	Davi Bezerra Macêdo Santos, Aliane Loureiro Krassmann and Leo Natan Paschoal
Abstract:	The relationship between companies and customers, particularly in service sectors such as insurance, is undergoing a significant transformation driven by the adoption of chatbots. Companies have widely implemented these systems as customer service support tools, aiming to enhance user experience through personalized and efficient solutions. However, despite their evident benefits, challenges related to understanding the nuances of language can lead chatbots to deliver inadequate, unsatisfactory, or artificial responses, which may frustrate users. Therefore, it is essential to examine the impact of chatbots on user experience, identifying both positive aspects and those that hinder interaction and user satisfaction. In this context, the present study evaluates the experience of insurance company clients when interacting with a customer service chatbot. For this study, clients of one of Brazil’s leading insurance companies were invited to use the company’s chatbot and complete the User Experience Questionnaire (UEQ), a tool designed to assess user experience. Following an analysis of the collected data, the results indicated that participants positively evaluated the system’s efficiency, highlighting its ability to provide fast and accurate responses. However, the need for improvements in personalizing interactions and enhancing emotional engagement was noted. Furthermore, the reliability analysis of the instrument revealed unsatisfactory internal consistency for one of the UEQ’s scales, raising questions about its suitability for evaluating user chatbot interactions.
Download

Paper Nr:	167
Title:	A Tool Supported by LLM for Heuristic Usability Evaluation
Authors:	Gabriel Felipe Jess Meira, Brunno Tatsuo Noda Bandeira da Cruz, Tiago Vieira Paulin, Pedro Henrique Dias Valle and Leo Natan Paschoal
Abstract:	Heuristic evaluation is one of the most established usability inspection methods in the field of Human-Computer Interaction; however, its effectiveness depends on the evaluator’s experience. Novice evaluators frequently face difficulties in the abstract interpretation of heuristics, leading to inconsistent findings. Although attempts have been made to automate the process using Large Language Models (LLMs), evidence suggests that full automation cannot replace expert human judgment. This paper presents PANDA, a web tool that integrates the heuristic inspection workflow with an LLM-based virtual assistant. The proposal’s key differentiator lies in its on-demand cognitive support, where the assistant acts as a real-time technical mentor, helping novices substantiate problems and classify severity. To evaluate the solution’s viability, an experimental study was conducted with 12 novice evaluators, comparing scenarios with and without the intelligent assistant’s support. The results, grounded in the Technology Acceptance Model (TAM), provide empirical evidence regarding perceived usefulness and ease of use, validating the tool’s potential to support the technical training of new inspectors.
Download

Paper Nr:	213
Title:	Diagnosing Socio-Technical Coordination and Interoperability Breakdowns through Service Design: A Case Study of an Airport Public Service Ecosystem
Authors:	Marcelo Judice, Andrea Judice, Glauco Pedrosa, Flavio Costa and Rejane Figueiredo
Abstract:	Socio-technical breakdowns emerge when misalignments between information systems, organizational processes, and stakeholder responsibilities undermine coordination and interoperability in complex service environments. In highly regulated contexts such as air transportation, these breakdowns frequently propagate along the passenger journey due to fragmented information flows and weak inter-organizational alignment. This study demonstrates how Service Design can be used as a diagnostic approach to systematically identify coordination and interoperability breakdowns within a complex airport public service ecosystem. We report a mixed-methods case study conducted at Brasília International Airport, combining ethnographic observations, semi-structured interviews with multiple stakeholders, participatory co-creation workshops, and an online survey with 189 passengers. The findings identify recurring breakdown patterns related to information synchronization across channels, fragmented service interaction protocols, weak coordination in accessibility assistance, limited interoperability in baggage handling processes, and insufficient real-time disruption management. These issues reflect misalignments across technical, semantic, and organizational dimensions of interoperability. By integrating Service Design methods with Information Systems (IS) perspectives on socio-technical coordination, this study provides empirical evidence of how user-centered, participatory approaches can support the systematic diagnosis of coordination failures in complex public services. The results contribute to IS research on coordination and interoperability and offer practical guidance for interoperability-aware service redesign and governance improvements in public-sector ecosystems.

Paper Nr:	222
Title:	Optimized EEG Channel Selection for Motor Imagery Using Quantum & Non-Dominated Sorting Genetic Algorithm
Authors:	Agampreet Singh, Arnav Dogra, Stuti Chug and Vijay Kumari
Abstract:	A brain–computer interface (BCI) acts as a bridge between the human brain and external systems, particularly for individuals who are unable to move their bodies. This connection is established through electrical signals recorded using an electroencephalogram (EEG). Although EEG signals can be acquired from up to 118 electrode locations, improper selection of electrodes may lead to redundant or irrelevant information for a specific task. To address this challenge, we propose a quantum-based genetic algorithm (QGA) integrated with the non-dominated sorting genetic algorithm II (NSGA-II) for efficient channel selection in motor imagery EEG classification. The QGA exploits quantum superposition to explore the search space more effectively, thereby reducing computational cost while preserving classification accuracy. The proposed algorithm is validated using the benchmark PhysioNet EEG Motor Movement/Imagery dataset. The proposed technique outperforms some existing methods and efficiently classifies motor imagery tasks.
Download

Paper Nr:	237
Title:	A Computerized Multimodal Battery for Emotion Recognition and Visual Attention Assessment in School-Aged Children
Authors:	Tiago Mota de Oliveira, Tatiele dos Santos Telaska, Claudemir Casa, Felipe Canarozzo Lourenco, Tatiana Izabele Jaworski de Sá Riechi and Luciano Silva
Abstract:	Emotion recognition assessment in children often relies on single-modality stimuli, which limits ecological validity and restricts the analysis of visual attention during task performance. This paper presents BACRE-I, a computerized multimodal battery designed to assess emotion recognition in school-aged children through six phases integrating static images, dynamic videos, vocal stimuli, audiovisual clips, avatar-based facial expressions, and eye tracking. The manuscript adopts a system-oriented perspective, emphasizing the architecture, multimodal organization, and phase-level behavior of the framework rather than subgroup neuropsychological comparisons. BACRE-I was experimentally applied to 82 children aged 8 to 12 years from Brazilian public schools. The results showed variation across stimulus modalities: the highest mean accuracy was observed in the vocal-stimulus phase (91.46%) and in the dynamic-video phase (82.18%), whereas the lowest mean accuracy was found in the avatar-based phase (45.93%). Eye-tracking records indicated that visual attention was concentrated mainly on central facial regions, especially the forehead, eyes, and nose, adding interpretive support to behavioral performance measures. These findings demonstrate the feasibility of BACRE-I as a structured computerized framework for collecting multimodal emotion-recognition data and suggest that combining behavioral and gaze-based measures can improve the interpretability of children’s responses across different task conditions.
Download

Paper Nr:	254
Title:	Distributed AI Orchestration in Web-of-Things for Accessible Indoor Assistance
Authors:	Seyed Shahabadin Nasabeh, Santiago Meliá, Jaume Aragonés, Barbara Leporini and Diana Gadzhimusieva
Abstract:	Indoor mobility remains a significant challenge for visually impaired people (VIPs), especially in complex environments. Fragmented assistive applications and limited contextual awareness reduce VIPs' autonomy and increase their cognitive load. Many existing solutions operate in isolation, failing to coordinate interaction across heterogeneous devices and services. This paper presents a Web of Things (WoT)-based information system with a distributed AI orchestrator that coordinates edge smartphones and cloud services to provide low-latency, context-aware indoor assistance. Built on the model-driven MoSIoT framework, this approach integrates six AI-based assistive capabilities, exposed as WoT services and activated by user intents, within a unified interaction architecture. The orchestrator mediates human-AI-environment interaction by dynamically allocating tasks across edge and cloud resources based on user profiles and accessibility requirements. This reduces interaction fragmentation and supports adaptive, context-aware interaction. We evaluated user experience and adoption in a cross-national formative study using the Technology Acceptance Model (TAM). The results show that perceived usefulness and perceived ease of use drive adoption, while perceived external constraints moderate behavioral intention. These results provide preliminary evidence that distributed AI orchestration can improve the perceived accessibility and usability of indoor assistive environments.
Download

Paper Nr:	280
Title:	Proposal and Validation of a Checklist for Persona Inspection in Software Development
Authors:	Enéas Mesquita Cunha Júnior, Williamson Silva and Ana Carolina Oran
Abstract:	Personas are widely used artifacts for representing archetypal users in software development, but the lack of standardized inspection methods often leads to artifacts based on unvalidated assumptions. This work proposes and validates InspecPersona, a structured checklist for persona inspection that integrates the Persona Perception Scale (PPS) with the PATHY template. The artifact was developed by integrating constructs from the PPS and the structural elements of PATHY, resulting in a checklist comprising 10 constructs and 33 items, organized into two parts: (1) persona structure and content, and (2) perceived value to the development team. An exploratory study with 41 Software Engineering students (organized in 8 project teams) produced 1,353 item responses. Results show strong agreement for Routine & Problems (≈95.2%), Empathy (≈95.9%), and Clarity (≈94.5%), while constructs such as Technology Use and Cognitive Preferences revealed gaps or neutrality. Acceptance measured via a TAM-based questionnaire was positive (Perceived Usefulness ≈83%, Perceived Ease of Use ≈80.5%, Behavioral Intention ≈73.2%), although Perceived Pleasure was low. Key limitations include ambiguous item wording and mismatch between some forced-choice statements and the agree/disagree response scale. As a contribution, InspecPersona is a validated instrument for evaluating personas in software projects and provides concrete directions for refining item formulation and response options in future studies.
Download

Paper Nr:	286
Title:	Design Thinking in Human-Centered Chatbot Engineering: A Systematic Mapping Study
Authors:	João Emilio Villa, Gilleanes T. A. Guedes, Paulo Silas Severo de Souza, Ricardo Vilela, Pedro Henrique Valle and Williamson Silva
Abstract:	Chatbot development has advanced considerably with the incorporation of modern Natural Language Processing (NLP) techniques. However, systematically integrating user-centered practices into chatbot engineering workflows remains an open challenge. Design Thinking (DT) has emerged as a promising approach to guide chatbot development. Nevertheless, its adoption in the chatbot development literature appears fragmented and is reported with varying degrees of methodological rigor. This paper presents a Systematic Mapping Study (SMS) that investigates how DT is applied in chatbot development across five dimensions: DT process models and phases, techniques used at each stage, application domains, implementation strategies, and evaluation practices. The results show that DT models are most commonly applied in healthcare, education, and organizational settings and are structured in five or six phases. In terms of techniques, personas, empathy maps, brainstorming, storyboards, and prototyping are among the most frequently used. We also observed a shift in implementation strategies from rule-based systems to NLP/NLU-based solutions and, more recently, to Large Language Model (LLM)-powered approaches. In addition, a substantial proportion of the selected studies report empirical evaluations of the resulting artifacts. This SMS organizes evidence on the use of DT in chatbot development, systematizes application and evaluation patterns, and highlights methodological gaps that should be addressed to strengthen empirical research in Human-Centered Software Engineering.
Download

Paper Nr:	99
Title:	An AI-Enhanced Empathy Map Technique: A Comparative Approach to Automated Semantic Context Extraction
Authors:	Nadia Menad, Mustapha Si-Tayeb, Roumaissa Farah and Imene Ait Mohamed
Abstract:	User Experience (UX) design requires a strong understanding of users, yet creating empathy maps from interviews and feedback can be time-consuming. This study proposes an AI-based tool to automatically generate empathy maps from textual user data. Two approaches were implemented: a basic Natural Language Processing (NLP) method and a hybrid approach combining NLP with the Naïve Bayes machine learning algorithm. A comparative analysis shows that the hybrid method improves the classification of user insights into empathy map elements. This approach supports faster analysis of user feedback and helps designers make more informed, user-centered decisions.

Paper Nr:	175
Title:	A Persona-Based WCAG 2.2 Mapping for Mobile Accessibility among Older Adults in Brazil
Authors:	Maíra Rocha Santos and Manuella Valadares
Abstract:	Population aging poses significant challenges to mobile accessibility, particularly for older adults with diverse functional, social, and behavioral characteristics. While accessibility standards such as the Web Content Accessibility Guidelines (WCAG) provide technical requirements, their translation into actionable design decisions remains difficult in early-stage mobile development. This study proposes an age-sensitive, persona-based decision-support approach to operationalize WCAG 2.2 for mobile interfaces used by older adults. Using secondary national datasets and qualitative evidence, two personas representing distinct age groups (60–79 and 80+) were constructed to capture health-related limitations, social contexts, and behavioral patterns. Identified accessibility barriers were systematically mapped to WCAG 2.2 Level A and AA success criteria, enabling traceable associations between user difficulties and standards-based solutions. Based on the severity of functional limitations observed in the 80+ group and the recurrence of barriers across personas, a compact, prioritized decision set was derived to support mobile design and heuristic evaluation. The results highlight the importance of legibility, contrast, touch error tolerance, navigation clarity, and media comprehension in ensuring basic access and autonomy for the oldest users. By translating age-sensitive user representations into concrete accessibility decisions, this study offers a lightweight decision-support artifact for designers and developers seeking to incorporate accessibility considerations into mobile interfaces for aging populations.
Download

Paper Nr:	224
Title:	Practitioner Perspectives: Usability Needs for Digital Sustainable Product Development Tools
Authors:	Yan Hu, Rachael K. Gould, Valeria Garro and Peng Wang
Abstract:	This study investigates the application of Nielsen’s 10 usability heuristics in the design of digital Sustainable Product Development (SPD) tools using a participatory design approach. Participants included representatives from three companies and SPD experts from academia. During the workshops, the ten usability criteria were presented, after which participants engaged in a brainstorming session to generate practical recommendations for SPD tool design based on their professional experience. The results suggest that SPD practitioners consider these heuristics relevant for SPD tool development. Future research will evaluate the effectiveness of integrating these guidelines into the design process and ongoing improvements to SPD tools.
Download

Area 6 - Enterprise Architecture

Full Papers

Paper Nr:	23
Title:	Challenges and Opportunities in Digital Transformation for Small and Medium-Sized Manufacturing Enterprises
Authors:	Alex Yiqiao Wang and Sherah Kurnia
Abstract:	Digital transformation is reshaping all sectors, with manufacturing small and medium-sized enterprises (SMEs) playing a crucial role in the sector’s value creation. Despite growing research interest in digital transformation, studies on manufacturing SMEs remain fragmented and often geographically limited. Through a systematic review of 16 of the most relevant studies across 11 countries and thematic synthesis, this study identifies nine key challenges and six opportunities. It further conceptualizes these findings from a socio-technical perspective to propose recommendations driving digital transformation. The study offers a modest contribution to the existing literature and provides actionable insights for advancing digital transformation in manufacturing SMEs.
Download

Paper Nr:	27
Title:	Mining Process Evolution in the Public Sector: A Dual-Perspective Case Study
Authors:	Rommel Roosevelt de Lima Sousa, Marcelo Fantinato and Sarajane Marques Peres
Abstract:	Over 11 years, a key administrative process in Brazil’s public sector shifted from a centralized, standardized structure to a distributed, adaptive configuration. This study examines this transformation within the SEI platform, addressing challenges of high variability and semantic ambiguity in broadly labeled tasks executable by any unit in varying sequences. Using a dual-perspective approach that combines control-flow process mining and organizational network analysis, we detected a shared concept drift across functional and organizational dimensions around the 6,000th case, which was used to segment the event log into two temporal phases. Process models for each phase were then built and evaluated using Inductive Miner, Split Miner, and X-Processes. The results show a transition from concentrated coordination and simpler execution paths to more distributed collaboration and structurally richer process behavior. Methodologically, the study offers a pragmatic drift-aware, dual-perspective protocol for analyzing high-variability public-sector workflows, showing how segmentation, variant filtering, and handoff-network analysis can support more interpretable and actionable process mining results.
Download

Paper Nr:	160
Title:	Rethinking Performance Indicator Research in BPM
Authors:	Leticia Naomi Asano, Encarna Sosa Sanchez and Lucineia Heloisa Thom
Abstract:	Since the discussion on Business Process Management began to gain relevance among researchers and business managers, the community has increasingly reflected on how performance indicators should be managed. The first proposals for indicators were quantitative, considering that the measures themselves are formulas and numbers used to make important decisions to drive organizational success. In this work, we are concerned with understanding how the performance indicator research has evolved over the last 63 years of research and whether a part of this evolution could be the formalization of the definition of such measures, rather than determining them through frameworks. This motivated a bibliometric analysis to drive qualitative research. We were able to understand that much of the research conducted nowadays still relies on frameworks built for the market of two decades ago, even though new results using formalization and other newer technologies, such as process mining, are being published by the community.
Download

Paper Nr:	306
Title:	Antifragile Principles in Enterprise Architectures
Authors:	Llanos Cuenca, Andrés Boza and Alix Vargas
Abstract:	Organizations increasingly operate in dynamic and uncertain environments characterized by rapid technological change and frequent disruptions. Traditional Enterprise Architecture (EA) approaches primarily focus on alignment, optimization, and stability, which may limit their effectiveness when organizations face unpredictable stressors. The ability of organizations to benefit from stress and disorder is called antifragility. This research explores how EA can evolve to better support organizations in such contexts by integrating antifragile principles. The study proposes an enhanced EA model that incorporates antifragile architectural principles aligned with the conditions of emergence: connectivity, diversity, rate of information flow, lack of inhibitors, good boundaries, intentionality, and anticipation. Architectural principles act as governance mechanisms that guide architectural design and decision-making. By defining antifragile architectural principles, organizations can incorporate key elements to address stressors such as technological disruption, market volatility, and organizational change. To illustrate the applicability of the proposed approach, the model is represented using the ArchiMate modelling language and demonstrated through a running example. The modelling highlights how antifragile principles can be embedded within architectural elements and relationships, providing a structured way to represent and analyse how enterprise architectures can adapt and evolve in response to disruptions.
Download

Paper Nr:	307
Title:	Integrating Prioritized Guidelines into BPMN Conformance Checking to Improve Model Clarity
Authors:	Thiago Richter, Manuela Campos de Amorim and Marcelo Fantinato
Abstract:	Business Process Model and Notation (BPMN) models often vary in clarity and quality, partly due to the large number of available modeling guidelines and the lack of systematic support for determining which recommendations are most critical. We assessed how six current BPMN modeling tools operationalize all forty modeling guidelines. We present an extension of BPMN Inspector that integrates guideline prioritization, obtained via the Analytic Hierarchy Process (AHP), into tool-supported guideline verification. The extension operationalizes a prioritized guideline list within the tool, provides prioritized visual feedback on guideline compliance, and introduces a weighted guideline-adherence score, the Understandability Index, together with its aggregated form across multiple models, the Overall Understandability Index. The proposed integration is intended to support issue triage by highlighting guideline violations according to the validated prioritization scheme.
Download

Short Papers

Paper Nr:	32
Title:	Stakeholder-Oriented Model Formatting
Authors:	Lovis Justin Immanuel Zenz, Peter Hillmann and Andreas Karcher
Abstract:	It is crucial to orient enterprise architecture models towards their stakeholders’ needs. In this paper, we assess formatting geared towards such stakeholder orientation. To this end, we collect requirements from existing research on information presentation, derive guidelines from these requirements, apply these guidelines to an exemplary set of initial models, and experimentally evaluate the added value for stakeholders. We conclude that appropriate application of formatting guidelines in modeling can indeed significantly improve the ease of information gathering. However, neglecting some requirements in favor of others can result in deterioration instead of the expected improvement. At the same time, additional factors should be considered when striving to meet stakeholder needs, such as the use of appropriate references – like enterprise architecture frameworks. Altogether, we furnish a universally applicable, concrete and measurable approach for orienting models towards their stakeholders that promotes ease of information gathering and inclusion through accessibility.
Download

Paper Nr:	33
Title:	Technology Capabilities Management Framework – TCMF
Authors:	Ralf Luis de Moura, Maria Stella Michirefe, Claudio Ramalho and Claudio Dal Col
Abstract:	The increasing complexity of hybrid digital ecosystems demands structured, strategy-aligned management of technological foundations. This article presents the Technical Capabilities Management Framework (TCMF), a model that formalizes technical capabilities as stable, technology-agnostic architectural constructs linking strategy, business capabilities, and technological execution. Drawing on dynamic capabilities theory, organizational semiotics, enterprise architecture standards, maturity models, and the digital transformation literature, the framework defines a metamodel, a lifecycle, a maturity assessment method, and a multi-horizon roadmap. A use case in a large industrial organization demonstrates how the TCMF clarifies capability definitions, exposes architectural gaps, and guides coherent technology planning. The findings show that the TCMF improves strategic alignment, strengthens architectural consistency, and supports systematic digital evolution, while highlighting implementation challenges.
Download

Paper Nr:	61
Title:	PM4SMEs: A Process Mining Methodology for Small and Medium-Sized Enterprises
Authors:	Rob Bemthuis
Abstract:	Process mining uses event data to discover, monitor, and improve business processes. Although it has demonstrated clear value in larger organizations, its use in small and medium-sized enterprises (SMEs) appears to be less mature. SMEs often operate under constraints such as limited analytics expertise, heterogeneous and loosely integrated digital tools, and sparse process documentation. As a result, process mining projects in these settings often begin with fragmented data and limited role specialization. Existing process mining methodologies provide useful guidance, but many assume more stable data access and more dedicated specialist capacity than SMEs can typically provide. In this paper, we present PM4SMEs, a process mining project methodology tailored to SME contexts. Following a design science research (DSR) approach, we derive design objectives from recurring SME challenges and instantiate them in an iterative, phase-based workflow. Each phase defines inputs, activities, outputs, roles, and lightweight deliverables intended to remain feasible for small teams. We illustrate the methodology through scenario-based walkthroughs that reflect common SME situations and data conditions, and we provide an initial formative assessment focused on analytical traceability and scenario-based illustration. Overall, PM4SMEs offers a structured yet pragmatic entry point for SMEs seeking to engage in process mining, while laying the groundwork for future empirical evaluation.
Download

Paper Nr:	78
Title:	Integrating Cybersecurity and Enterprise Architecture to Improve Cyber Risk Management: A Practitioner’s View
Authors:	Nick Nieuwenhuis, Edzo Botjes, Raymond Slot, Christian-Alexander Bunge and Martine Groen
Abstract:	Integrating cybersecurity with enterprise architecture (EA) has been proposed to provide a more holistic, proactive, and structured approach to managing cyber risks, but practical guidance on how to achieve it is limited. To address this, we conducted a literature review identifying 17 peer-reviewed articles that articulate enterprise-level strategies for integrating cybersecurity and EA. We validated and analysed these findings with cybersecurity and EA practitioners from Dutch enterprises. Using thematic analysis, we identified six key integration strategies: (1) Integrating cybersecurity into EA framework, (2) adopting the ‘Security by Design’ principle, (3) integrating business requirements with security requirements, (4) leveraging EA to provide input for cyber risk assessment, (5) mapping identified cyber risks to EA assets, and (6) aligning business & IT activities. In addition, we have identified key enablers and barriers which impact the successful implementation of EA and cyber risk management. In conclusion, we show that cybersecurity and EA integration positively influence the enterprise’s cyber risk management, ensuring traceability between business objectives and security controls.
Download

Paper Nr:	97
Title:	The Conceptualisation of Lean Enterprise Architecture: A Systematic Synthesis for Agile IT Operations
Authors:	Songzi Xu, Rod Dilnutt, Leila Meratian Esfahani and A. B. M. Nayeem
Abstract:	Organisations face increasing pressure to improve the agility and efficiency of their information technology (IT) operations while facing significant financial, organisational, and technical barriers. Traditional Enterprise Architecture (EA), though widely recognised for aligning business and IT, is frequently criticised for rigidity, complexity, and limited value realisation in fast-changing environments. In response, Lean Enterprise Architecture (Lean EA) has emerged as an alternative that emphasises value orientation, minimalism, and adaptability. However, its conceptual foundations remain fragmented across existing studies. This paper presents a systematic review of existing Lean EA studies to clarify its conceptual foundations. Through qualitative thematic analysis, four core principles of Lean EA are defined: Value-Driven and Business-Outcome Focus, Minimalist and Just-in-Time Approach, Process Efficiency and Waste Elimination, and Continuous Adaptability. This study presents the first coherent conceptualisation of Lean EA, explaining how architectural practices can support agile and value-focused IT operations. The findings contribute to EA research by consolidating dispersed insights into a structured set of principles to provide a conceptual reference for practitioners to reflect on and enhance architectural practices in dynamic digital environments.
Download

Paper Nr:	107
Title:	Tax Compliance Management Systems for SMEs: A Functional and Architectural Analysis
Authors:	Viet Anh Do and Rainer Alt
Abstract:	Tax Compliance Management Systems (TCMS) form a domain-specific subsystem within an organization’s Enterprise Architecture (EA). However, small and medium-sized enterprises (SMEs) often lack the structures needed to integrate tax-relevant processes, documentation, and controls in a structured and auditable manner conforming to IDW PH 1/2016. This paper examines how TCMS functionalities can be designed and embedded within the EA of SMEs. Based on a structured derivation of TCMS functional areas and an evaluation of fifty software products, four core areas are identified: Documentation, Reporting, Communication and Coordination, and Process Management and Automation. A market analysis supported by principal components analysis (PCA) groups existing applications into three clusters-integrated compliance suites, automation and integration platforms, and analytics-collaboration backbones-which serve as software-architectural patterns for configuring modular TCMS solutions. Building on these, the paper develops options that align TCMS requirements with available system types and integration mechanisms typical of SMEs. The results show that TCMS implementation in SMEs constitutes a design task in which functional priorities must be matched with suitable software-architectural patterns to operationalize an SME-compatible TCMS configuration.
Download

Paper Nr:	117
Title:	A Semantic Microservice Middleware Architecture for Integrating BIM, ERP, and IoT in Construction
Authors:	Elizaveta Stashevskaia, Thomas Sepanosian, Omar Mohamed Fawzy, Rob Bemthuis and Simon Hacks
Abstract:	Construction IT landscapes are fragmented across Building Information Modeling (BIM), Enterprise Resource Planning (ERP), IoT platforms, and logistics systems. This fragmentation limits data reuse and makes coordinated decision making challenging. We propose a microservices-based semantic middleware architecture that integrates heterogeneous systems through ontology-informed service contracts. The approach combines API-first decomposition for incremental adoption with ontology-informed mediation. It uses IFC (built-environment concepts) and SAREF (IoT semantics) to support consistent interpretation across sources. A prototype shows structured cross-system exchanges via REST/OData endpoints. It also illustrates how semantic annotations can be attached to payloads for downstream use. We evaluate the approach through quality-attribute reasoning, expert discussions with industry stakeholders, and a migration plan. The work provides a practical path for modernizing construction integration without replacing existing systems.
Download

Paper Nr:	210
Title:	Interaction Model for Enterprise Information Environments with Large Language Models
Authors:	Ekaterina Mashina and Florian Wahl
Abstract:	In the age of pervasive uncertainty and increasingly powerful intelligent models, organizations face growing challenges in aligning probabilistic reasoning with structured enterprise processes. This article proposes a comprehensive conceptual model for integrating large language models (LLMs) into the Enterprise Information Environment (EIE) to address the fragmentation present in existing solutions. The scientific contribution of the work lies in structuring the relationship between the probabilistic nature of LLM generation and the deterministic algorithms of corporate information systems. The innovative contribution consists in the proposal of a conceptual service-oriented architectural model for LLMs and the EIE, encompassing semantic configuration, data integration, information storage and disposal, as well as interpretation and integration processes. In contrast to existing highly specialized models, the work adopts a comprehensive perspective on LLM–EIE interaction, thereby providing a conceptual foundation and common vocabulary to guide future engineering and governance for management and critical business processes. The work enables synchronization between LLM cognitive capabilities and formalized corporate knowledge. This synchronization mitigates risks associated with the “black box” problem and uncertainty in generated results.
Download

Paper Nr:	236
Title:	Exploring Behavioral and Psychological Factors Influencing Customer Engagement in Saudi Airline e-Commerce for Future CLV Prediction
Authors:	Ayat Baroom, Kawther Saeedi and Bahjat Fakieh
Abstract:	This study investigates behavioral, emotional, and platform related factors that shape customer engagement in Saudi airline operators as a foundation for future Customer Lifetime Value (CLV) prediction. A total of 524 valid responses were collected and analyzed using descriptive statistics to provide an initial understanding of customer patterns and engagement behaviors. The findings highlight dominant demographic groups, preferred booking platforms, travel purposes, browsing and booking frequencies, discount sensitivity, and varying levels of impulse driven behavior. Moreover, reliability and construct level correlation analyses were conducted to assess the internal consistency of behavioral measures and explore relationships among impulse tendencies, emotional triggers, and purchase behavior. Results indicate strong reliability for impulse buying tendencies (α = 0.832), acceptable reliability for emotional triggers (α = 0.753), and modest but theoretically expected reliability for purchase behavior (α = 0.607). The correlation matrix further revealed meaningful associations among the three behavioral constructs, supporting their relevance for future CLV related behavioral modeling. These insights form the foundation for future work, which will extend the framework using advanced analytical techniques.
Download

Paper Nr:	256
Title:	Software Modernization: Phase-Specific ICT Requirements, Governance and Resilience Alignment
Authors:	Ada Slupczynski and Beāte Krauze
Abstract:	Software modernization is increasingly shaped by evolving EU information and communication technology (ICT) legislation alongside rapid technological change. Yet many organizations still treat compliance as a post-deployment task, leading to additional costs, delivery delays, and Enterprise architecture (EA) debt. This paper argues that operational compliance feasibility should be treated as mandatory best practice for EA and integrated from the earliest modernization phases because many obligations translate into continuous operational activities. Based on an exploratory analysis of five EU legal frameworks related to ICT, we identify nine compliance aspects and map them to the modernization phases proposed by Warren and Ransom. The mapping shows that the deployment phase is involved in eight of nine aspects (dominant in three and supporting in five), that planning dominates with four aspects and that cross-phase dependencies require coordinated governance across the entire life cycle. These findings challenge the existing approach and indicate that ICT requirements must drive upfront EA decisions, requiring distinct expertise, budget allocation and governance mechanisms per phase.
Download

Paper Nr:	173
Title:	Adoption of Low-Code Development Platforms in the Public Sector: A Systematic Literature Review
Authors:	Eduardo de Oliveira Castro and Rejane Maria da Costa Figueiredo
Abstract:	Digital transformation has become a strategic priority for public sector organizations aiming to improve service delivery, efficiency and responsiveness under increasing resource and regulatory constraints. Low-Code Development Platforms (LCDPs) seem to be one alternative to speed up the digitalization and automation of government services and reduce dependence on traditional software development approaches. This paper presents a systematic literature review (SLR) on organizational, technical and economic factors related to the adoption of LCDPs in the public sector. The findings indicate that LCDPs might provide faster service delivery, empower business teams, foster innovation and reduce costs, but also pose risks related to governance, interoperability with legacy systems and information security. The primary contribution of this study is an integrative synthesis that situates LCDP adoption in the public sector not merely as a technical decision but as a socio-technical and organizational transformation. The results will contribute to the scientific literature and inform public organizations engaging in low-code to make evidence-informed decisions.
Download

Paper Nr:	208
Title:	Architectural Patterns for Data Mesh and Data Space Implementations
Authors:	Attila Papp and Udo Bub
Abstract:	Data-driven decision-making requires scalable, reliable data sharing, yet organizations face bottlenecks in centralized platforms and sovereignty constraints in cross-organizational exchange. This paper compares the socio-technical paradigms of Data Mesh (intra-organizational) and Data Spaces (inter-organizational) from a technical perspective, a pattern-based lens grounded in Design Science Research. It identifies two complementary Data Mesh patterns: logically decentralized meshes, in which domains own data products on a shared platform, and physically decentralized meshes, in which domains operate separate environments and federate discovery and access. For Data Spaces, it characterizes implementations as compositions of recurring patterns, notably connector-centric peer-to-peer exchange and explicit separation of the control plane from the data plane. The paper analyzes catalogs as a convergence point, showing how their characteristics differ across paradigms and how governance is executed, and synthesizes four delivery patterns (in-place access, replication, streaming, compute-to-data) to clarify trade-offs in sovereignty, the feasibility of technical controls, and operational complexity.
Download

Paper Nr:	229
Title:	Managing Model Complexity through Structure
Authors:	Lovis Justin Immanuel Zenz, Harald Hagel, Peter Hillmann and Andreas Karcher
Abstract:	Structuring models is crucial for ensuring ease of gathering contained information. While enterprise architecture frameworks and other references inspire some structure, they often fall short when models grow too large. Hence, managing such complexity necessitates further structuring. In this paper, we present a four-step systematic process for sustainably achieving such structuring in real-world scenarios. Our approach ensures stakeholder and perspective orientation, maintains compliance with stipulations and preserves any preexisting structure. Altogether, it balances structural rigor with long-term practical applicability.
Download

Paper Nr:	245
Title:	A Process Mining Case Study in IT Incident Management
Authors:	Oscar Avila, Mathias Weske and Alexander Estacio
Abstract:	This article presents a case study on the application of process mining techniques in IT incident management within the Department of Information and Technology Services (DSIT) at a university in Colombia. The study aims to bridge the gap on the use of process mining for diagnostics in IT service management (ITSM) projects. The methodology involves generating an event log from 3440 incident registers exported from the CRM platform this institution uses to manage IT incidents, followed by process discovery and attribute-based segmentation to identify patterns in process execution. The case study demonstrates the feasibility of applying process mining techniques to the diagnostics of IT management processes, highlighting challenges encountered and solutions applied. The results indicate that process mining can help align IT management processes with performance goals, providing valuable insights for the DSIT’s ongoing transformation project.
Download

Paper Nr:	275
Title:	Teleophthalmology During Crisis: An Experience of Burkina Faso’s Army
Authors:	Wendgounda Francis Ouedraogo and François Nadziga
Abstract:	Teleophthalmology is increasingly establishing itself as an essential tool for eye health in various crisis situations, with the recent COVID-19 pandemic serving as a notable example. In the security crisis that Burkina Faso has been facing since 2015, the need to treat soldiers with eye injuries on the front lines has led to the implementation of this healthcare practice. The case study of ophthalmological teleconsultations in the Burkinabè military health service shows encouraging results. While this practice has remained feasible with the limited resources available, a completely different approach incorporating smart solutions is necessary. A knowledge-based system, for example, could effectively meet expectations (diagnostic support, continuous learning) if the requirements in terms of design and deployment are aligned with the local context.
Download

Paper Nr:	303
Title:	Governing Emissions Data: Institutional Perspectives on Data Governance in Agricultural Supply Chains
Authors:	Alex Yiqiao Wang and Sherah Kurnia
Abstract:	Emissions data governance in agricultural supply chains requires data to be captured at the point of origin, managed consistently within organizational boundaries, and shared transparently across supply chain partners. In practice, these three requirements remain misaligned, yet existing data governance frameworks, designed primarily from an intra-organizational perspective, are ill-equipped to address this misalignment. Existing studies further tend to focus on how emissions data can be shared rather than why current practices take the form they do, leaving the institutional forces that sustain them unaddressed. Drawing on new institutionalism and institutional entrepreneurship, this study develops a theoretical foundation for examining how emissions data capture, management, and sharing practices form, persist, and are reshaped across agricultural supply chains. Three progressive research questions are proposed, to be addressed through an engaged scholarship methodology in collaboration with industry partners. This proposed study aims to contribute theoretical understanding of the institutional dynamics shaping emissions data practices, a theory-informed reconfiguration of data governance mechanisms suited to inter-organizational settings, and actionable guidance for practitioners.
Download