DSpace Repository

Now showing 1 - 5 of 5

Trabalhos acadêmicos e de conclusão de curso

Recent Submissions

TRANSFORMAÇÃO RESILIENTE PARA DETECÇÃO DE ANOMALIAS EM SÉRIES TEMPORAIS HETEROSCEDÁSTICAS
(CENTRO FEDERAL DE EDUCAÇÃO TECNOLÓGICA CELSO SUCKOW DA FONSECA, 29/01/2026) LAIS BARONI VILLET; EDUARDO SOARES OGASAWARA; RAFAELLI DE CARVALHO COUTINHO; EDUARDO SOARES OGASAWARA; RAFAELLI DE CARVALHO COUTINHO; GUSTAVO PAIVA GUEDES E SILVA; FABIO ANDRE MACHADO PORTO; CARLOS EDUARDO RIBEIRO DE MELLO; EDUARDO FONSECA MENDES
Heteroskedastic time series exhibit non-constant variance over time, which compromises the effectiveness of classical anomaly detection methods based on fixed thresholds. In particular, this characteristic affects the detection of contextual anomalies, whose detection relies on local dispersion. To address this challenge, this paper proposes a transformation technique called Resilient Transformation (RT), which modifies the original series through empirical mode decomposition, roughness-based high-frequency component selection, differencing, and local dispersion normalization. This transformation equalizes variance and highlights anomalies. Evaluations using the Yahoo! S5 dataset, within the Harbinger framework, show that RT improves the performance of classical methods and can be combined with thresholding to form a standalone detection method called Resilient Transformation Anomaly Detector (RTAD).
RECONHECIMENTO DE ENTIDADES NOMEADAS EM RELATOS CRIMINAIS INFORMAIS COM APOIO DE METADADOS ESTRUTURADOS
(CENTRO FEDERAL DE EDUCAÇÃO TECNOLÓGICA CELSO SUCKOW DA FONSECA, 01/07/2026) GUSTAVO MELO; EDUARDO BEZERRA DA SILVA; KARLA TEREZA FIGUEIREDO LEITE; EDUARDO BEZERRA DA SILVA; KARLA TEREZA FIGUEIREDO LEITE; RONALDO RIBEIRO GOLDSCHMIDT; GUSTAVO PAIVA GUEDES E SILVA; KELE TEIXEIRA BELLOZE
This work investigates the problem of Named Entity Recognition (NER) in informal criminal reports registered through the Disque Denúncia service. These reports, often characterized by colloquial language, spelling errors, and free-text structure, pose significant challenges to the application of traditional Natural Language Processing models. In addition to free text, the reports are accompanied by structured metadata, such as occurrence type, location, and date, which may provide additional contextual information relevant to the task. In this study, an approach based on fine-tuning large language models is proposed, using a manually annotated corpus, annotated by four different individuals and the dissertation author, containing entities of type Person, Location, and Organization. To address the scarcity of labeled data, the methodology includes the application of pseudo-labeling on a second and significantly larger corpus, thus expanding the training set in a semi-supervised manner. Additionally, report metadata are incorporated as a contextual source both during preprocessing and throughout the evaluation and refinement stages of the models. The experiments were conducted using the GliNER model and evaluate how the use of metadata and pseudo-labeling can contribute to improving model performance in informal corpora, with positive impacts on automated information extraction in public security contexts. The results reinforce the potential of hybrid and domain-sensitive approaches for real-world Natural Language Processing applications in environments with limited annotated data.
ESTUDO DA CARDINALIDADE E DA RELAÇÃO ENTRE ATRIBUTOS NA IMPUTAÇÃO DE DADOS CATEGÓRICOS
(CENTRO FEDERAL DE EDUCAÇÃO TECNOLÓGICA CELSO SUCKOW DA FONSECA, 01/10/2026) FELIPE DE AVILA TAVARES; JORGE DE ABREU SOARES; JORGE DE ABREU SOARES; DIEGO NUNES BRANDAO; RONALDO RIBEIRO GOLDSCHMIDT
This dissertation investigates the problem of categorical data imputation, focusing on the comparative analysis of different imputation methods and on how dataset structural characteristics influence their performance. The study considers three distinct scenarios: simple datasets composed predominantly of numerical attributes, mixed datasets containing both numerical and categorical variables, and fully categorical datasets. The proposed methodology artificially generates missing values under the Missing Completely at Random (MCAR) mechanism at rates of 10%, 20%, and 30%, using multiple random seeds to evaluate the stability of the analyzed methods. Approaches based on KNN, Multi-Layer Perceptron (MLP), and Random Forest were investigated, considering both imputation accuracy and the performance of classifiers trained on the imputed data. In addition, Cramer’s V coefficient was used to analyze the influence of correlation between categorical attributes in the imputation process. The obtained results indicate that global model-based methods, especially Random Forest, achieve greater robustness and stability against increasing missingness rates and dataset structural complexity, while KNN shows greater sensitivity to data heterogeneity and information loss. The findings reinforce the importance of considering characteristics such as cardinality and attribute correlation when selecting the most appropriate imputation method.
DETECÇÃO DE MOTIFS EM SÉRIES TEMPORAIS USANDO AUTOENCODERS: UMA ANÁLISE DO USO EM DADOS DE ECG
(CENTRO FEDERAL DE EDUCAÇÃO TECNOLÓGICA CELSO SUCKOW DA FONSECA, 15/06/2026) LUIZ CLAUDIO LEMOS DE OLIVEIRA; EDUARDO SOARES OGASAWARA; EDUARDO SOARES OGASAWARA; LAURA SILVA DE ASSIS; HELGA DOLORICO BALBI; REBECCA PONTES SALLES
The discovery of motifs in biomedical time series, such as electrocardiograms (ECG), consists in identifying recurrent patterns with diagnostic potential. Traditional methods such as SAX present limitations due to their dependence on rigid statistical assumptions, which are inadequate for complex physiological signals. In parallel, autoencoders demonstrate superior capacity for learning nonlinear representations, but their application to motif discovery in ECG remains unexplored — a gap that this work addresses. This work proposes LEMOD, a method whose central contribution is the replacement of the statistical discretization of SAX by adaptive neural encoding via autoencoders, combined with a mechanism for searching and ranking motifs by frequency of occurrence and Shannon entropy. The methodology comprises three stages: (i) validation of the autoencoder’s reconstruction capacity; (ii) model training; and (iii) detection, parametric sensitivity analysis and generalization evaluation, with systematic comparison to SAX. Experiments were conducted using data from the MIT-BIH Arrhythmia Database. In the validation stage, reconstruction fidelity above 99% was observed. In detection tests, LEMOD achieved a relative gain of 92% in recall and 78% superiority in F1 over SAX, evidencing greater sensitivity to subtle morphological patterns without compromising specificity. The average entropy of the motifs (0.44 bits versus 0.90 bits for SAX) was lower, indicating greater internal morphological homogeneity — evidence that neural encoding identifies more coherent recurrent structures than traditional discretization. Detection time was 15% lower in LEMOD. The cross-detection experiment confirmed generalization of the learned latent representations to distinct series of the same lead. It is concluded that replacing traditional discretization with adaptive neural encoding is viable and provides quantitative and qualitative gains in motif discovery in ECG signals, establishing a methodological basis for automated biomedical signal analysis tools.
ANÁLISE DE DESEMPENHO E CARACTERÍSTICAS DE CORREDORES PARA PREDIÇÃO DE RESULTADOS E GERAÇÃO DE FEEDBACKS PERSONALIZADOS
(CENTRO FEDERAL DE EDUCAÇÃO TECNOLÓGICA CELSO SUCKOW DA FONSECA, 27/04/2026) NATHALIA CARVALHO TITO; GLAUCO FIOROTT AMORIM; EDUARDO BEZERRA DA SILVA; GLAUCO FIOROTT AMORIM; EDUARDO BEZERRA DA SILVA; DIEGO NUNES BRANDAO; CLAUDIO MICELI DE FARIAS
The increasing number of recreational runners has intensified the demand for solutions capable of providing individualized training support, especially among amateur athletes who often lack continuous professional guidance. In this context, this study proposes an integrated performance analysis model based on machine learning techniques, aiming to understand training patterns, predict runners’ performance, and generate personalized recommendations based on controllable variables. Data were obtained from a questionnaire and activity records exported from a running application, involving 26 athletes with different levels of experience, allowing a multidimensional view of training habits and sports experience. The methodology was structured in three main phases. In the first phase, a clustering analysis was performed using classical algorithms (K-means, DBSCAN, and Agglomerative Clustering) applied to principal components explaining 80% of the data variability. The selected model identified three distinct profiles: “Young Experienced,” “Less Experienced,” and “Veteran Experienced,” differentiated by age, sports maturity, and training patterns. In the second phase, a predictive model based on gradient boosting was developed using the XGBoost algorithm, both in a general configuration and in versions specific to each cluster. Linear regression models were also tested as a reference approach; however, XGBoost achieved superior performance across all evaluated scenarios, demonstrating a greater ability to capture nonlinear relationships and complex interactions among variables. The results also indicated that each group responded differently to training variables, reinforcing that segmentation substantially improves predictive performance and the adequacy of personalized recommendations. Model interpretability was investigated using SHAP values, which enabled the identification of the most influential variables in the predictions. In general, variables directly related to training structure and performance stood out, such as minimum and maximum speed, distance covered, pace variability, and recent performance history, as well as contextual factors such as temperature. The segmented analysis revealed distinct patterns across groups, indicating that different runner profiles present specific sensitivities to aspects such as intensity, regularity, and training consistency, further supporting the importance of personalized recommendations. In the third phase, an optimization algorithm was applied to identify combinations of controllable variables capable of maximizing the predicted performance for each profile. The evaluation of results was based on comparisons with baseline scenarios, in which decision variables were not adjusted, allowing objective quantification of the gains achieved through optimization. The resulting recommendations showed internal coherence and alignment with cluster characteristics: greater emphasis on intensity and training variety for young experienced runners; focus on regularity, consistency, and strength training for less experienced runners; and strategies centered on maintenance, balance, and load control for veteran runners. These findings demonstrate that the integration of clustering, predictive modeling, and optimization provides a consistent and promising approach for developing data-driven intelligent sports recommendation systems. Despite limitations related to sample size and the absence of more granular physiological indicators, the study provides initial evidence that computational models can effectively support personalized training in an efficient, accessible, and scalable manner. Future research may expand the dataset, incorporate additional informational dimensions, validate the model in larger populations, and explore its implementation in digital platforms. In conclusion, the combination of data science techniques and optimization methods significantly contributes to understanding running performance and to developing individualized recommendations that promote improvement, adherence, and safety in sports practice.