https://repositorio.ufpb.br/jspui/handle/tede/5439 PPGI Fri, 30 Oct 2026 03:17:16 GMT 2026-10-30T03:17:16Z http://repositorio.ufpb.br:443/jspui/retrieve/30178/ https://repositorio.ufpb.br/jspui/handle/tede/5439 https://repositorio.ufpb.br/jspui/handle/123456789/39398 Título: Classificação de consultas fiscais na SEFAZ-PB: uma análise comparativa entre método estatístico e modelos baseados em transformers Autor(es): Carvalho, Sarah Andrade Toscano de Orientador: Rêgo, Thaís Gaudencio do Abstract: Automatic classification of tax queries submitted to public digital systems, such as SefazVirtual from the State Department of Finance of Para´ıba (SEFAZ-PB), involves natural language messages sent by taxpayers through institutional service channels. These queries express doubts, questions, or requests related to taxes, legislation, and administrative procedures. The main challenge lies in the thematic overlap between categories, linguistic ambiguity, and the scarcity of annotated data for supervised training. Therefore, this research investigates and compares different modeling paradigms to mitigate these difficulties, including a classical statistical baseline based on TF-IDF combined with Logistic Regression, supervised fine-tuning of BERT (Bidirectional Encoder Representations from Transformers), and in-context learning strategies using the LLaMA 3.3 70B model. The dataset, composed of messages distributed across five categories related to the Tax on the Circulation of Goods and Services (ICMS) and the Motor Vehicle Property Tax (IPVA), underwent preprocessing steps including anonymization in compliance with the Brazilian General Data Protection Law (LGPD), text normalization, tokenization, and stopword removal. Two experimental training configurations were conducted for BERT (base and large), each evaluated using two versions of the dataset — the original version and an enriched version obtained through controlled rewriting of the queries with the support of ChatGPT and the expansion of label definitions. The TF-IDF model combined with Logistic Regression was used as a baseline to contextualize the performance gains of neural architectures. The evaluation, based on accuracy, precision, recall, and F1-score, showed that data augmentation increased the accuracy of BERT-large from 0.62 to 0.77 and BERT-base from 0.62 to 0.66, while the TF-IDF baseline combined with Logistic Regression achieved an accuracy of 0.80 with enriched data, and LLaMA 70B reached up to 0.87 accuracy in a few-shot scenario with 20 examples. These results confirm the viability of both statistical approaches and transformer-based models adapted to the tax domain as effective alternatives for optimizing the routing of queries and reducing delays in public service responses. Editor: Universidade Federal da Paraíba Tipo: Dissertação Wed, 25 Feb 2026 00:00:00 GMT https://repositorio.ufpb.br/jspui/handle/123456789/39398 2026-02-25T00:00:00Z https://repositorio.ufpb.br/jspui/handle/123456789/39380 Título: Modelo Dinâmico de Análise de Sobrevivência Aplicado à Previsão do Tempo de Internação em Hospitais Públicos Municipais Autor(es): Crispim, Gizeuda Cariry Lacet de Barros Orientador: Souza Filho, Guido Lemos Abstract: Healthcare service overcrowding poses significant challenges to hospital management, with Length of Stay (LOS) being one of the main performance and planning indicators. In this context, this study proposes a dynamic survival analysis approach to predict hospital length of stay, incorporating temporal updating of risk estimates throughout the patient’s hospitalization. The proposed framework is evaluated using real-world data from the public hospital system of João Pessoa, Brazil, aiming to provide predictive estimates that better reflect the evolving nature of discharge risk over time. As part of the validation process, the proposed dynamic approach is assessed in comparison with well-established survival models, including the Cox Proportional Hazards model (CoxPH), Random Survival Forest, and DeepSurv. The study encompasses fundamental concepts of machine learning and deep learning, data preprocessing techniques, and survival analysis theory, structuring model development within a classical machine learning pipeline. The main outcome of this research is a predictive model capable of estimating conditional and dynamically updated probabilities associated with hospital length of stay, allowing recalibration as new clinical information becomes available. Model performance is evaluated using the concordance index (C-index), enabling comparison between the proposed dynamic framework and traditional approaches. Results indicate that dynamic modeling has the potential to enhance predictive accuracy and practical applicability, reinforcing the feasibility of artificial intelligence techniques in developing intelligent clinical decision-support systems and improving hospital bed management efficiency. Editor: Universidade Federal da Paraíba Tipo: Dissertação Fri, 27 Feb 2026 00:00:00 GMT https://repositorio.ufpb.br/jspui/handle/123456789/39380 2026-02-27T00:00:00Z https://repositorio.ufpb.br/jspui/handle/123456789/39241 Título: Otimização em nível de bloco da ResNet com algoritmo genético para classificação de imagens endoscópicas Autor(es): Souza Junior, Gilberto Rodrigues de Orientador: Cabral, Lucídio dos Anjos Formiga Abstract: Convolutional Neural Networks have demonstrated capability in identifying patterns in medical images, however, the need for explainable models remains critical in high-risk medical applications. This study presents an optimization framework based on Neural Ar- chitecture Search using a Genetic Algorithm to improve the performance of ResNet-152 and ResNet-50 on the Kvasir v2 dataset. The proposed approach incorporates image preproces- sing techniques and selective block optimization, achieving an accuracy of 97.75% on the validation set and 97.12% on the test set. In addition, the analysis of the optimization process revealed evidence of the structural relevance of specific bottleneck blocks, identified as being sensitive to dataset characteristics. The results of this study contribute to advan- cing global explainability in ResNet-based solutions by providing a deeper understanding of model behavior when applied to endoscopic images. Editor: Universidade Federal da Paraíba Tipo: Dissertação Wed, 11 Feb 2026 00:00:00 GMT https://repositorio.ufpb.br/jspui/handle/123456789/39241 2026-02-11T00:00:00Z https://repositorio.ufpb.br/jspui/handle/123456789/39189 Título: Detecção de falhas em unidades geradoras de usinas hidrelétricas utilizando DCT e simulações Bootstrap Autor(es): Sartor, Matheus Lira Orientador: Souza, Leandro Carlos de Abstract: Fault Detection in generating units is a matter of utmost importance for maintenance in hydroelectric power plants, as anticipating faults can mean avoiding losses in energy supply to the national system, as well as potential savings through preventive actions or component replacement. The main challenge for fault detection is that not every anomaly data necessarily indicates a fault, so the model to be used must be capable of identifying intrinsic data characteristics and detecting an anomaly only when a specific fault signature exists. In an effort to address this issue, the DCT-Bootstrap model was developed and tested in real scenario of a power plant and on a reference dataset, successfully outperforming state-of-the-art reference models in both situations. This work thus introduces a new model that can be employed in fault detection tasks across industry scope. Editor: Universidade Federal da Paraíba Tipo: Dissertação Tue, 20 Feb 2024 00:00:00 GMT https://repositorio.ufpb.br/jspui/handle/123456789/39189 2024-02-20T00:00:00Z