https://repositorio.ufu.br/handle/123456789/18920 2026-09-17T20:08:21Z https://repositorio.ufu.br/handle/123456789/50172 Title: Explorando o uso de árvores B+ na indexação de dados por similaridade Abstract: Similarity queries are beneficial for retrieving complex data (such as multimedia data: images, videos, and audios), for which order relationships are not significant. In many application domains, ordering results based on distance calculations makes the data recovery process more intuitive. The optimization of operations involving similarity calculations are usually given by indexing data on methods, called Metric Access Methods (MAM). Despite the development that occurred over the last two decades, the presence of a large number of dimensions in the data degrades the performance of existing methods. In this context, this dissertation presents a new metric access method for similarity queries, such as range queries and nearest neighbor queries, through adding reference pivots to structures called B + trees. Experimental results performed with different real datasets demonstrate the effectiveness of the presented MAM, called GroupSim+. 2019-11-29T00:00:00Z https://repositorio.ufu.br/handle/123456789/50157 Title: Segmentação semântica de núcleos em leucoplasia bucal por meio de comitês de modelos de aprendizagem profunda Abstract: The histopathological assessment of oral potentially malignant disorders, particularly oral leukoplakia, relies heavily on the analysis of nuclear morphology. However, this process is marked by subjectivity and substantial interobserver variability. In this context, accurate cell nuclei segmentation is a fundamental step for computer-aided histopathological analysis systems. Despite its relevance, nuclei segmentation specifically in oral leukoplakia images remains underexplored, as does the use of ensembles built from heterogeneous architectures. This dissertation proposes and evaluates ensembles of heterogeneous deep learning models for the semantic segmentation of cell nuclei in histopathological images of oral leukoplakia. Ten architectures representing three paradigms were investigated: convolutional, hybrid, and attention-based. The models were evaluated individually and combined using five \textit{ensemble learning} strategies under different data augmentation configurations. The evaluation employed traditional segmentation metrics and statistical validation. As a methodological contribution, three individual-contribution metrics were proposed and formalized in this dissertation: \textit{kept rate}, \textit{discarded rate}, and \textit{error mitigation}. These metrics quantify, respectively, the preservation of an individual model's correct predictions, the loss of those correct predictions after combination, and the correction of individual errors by the ensemble. Among the individual models, the best performance was achieved by ResNeSt-Unet++, with a Dice coefficient of 0.8853. Combination strategies based on probabilistic averaging proved to be the most effective, with \textit{soft voting} achieving a Dice coefficient of 0.8924 and statistically significant gains over the individual models. The application of the proposed metrics indicated that architectures from distinct paradigms exhibit complementary error patterns, favoring the preservation of correct predictions and the mitigation of individual errors. In addition, the gain provided by the ensembles tended to saturate when three to four models were combined. These results demonstrate that heterogeneous ensembles constitute an effective strategy for nuclei segmentation in oral leukoplakia. 2026-08-11T00:00:00Z https://repositorio.ufu.br/handle/123456789/50148 Title: Avaliação do uso de grandes modelos de linguagem na análise de riscos em segurança da informação: um estudo comparativo com especialistas humanos Abstract: The increasing sophistication of cyber threats and the continuous evolution of the technological landscape impose significant challenges on organizations, especially given the global shortage of qualified cybersecurity professionals. In this context, Generative Artificial Intelligence and Large Language Models emerge as promising alternatives to support cybersecurity risk analysis and assessment activities. Despite their potential, Large Language Models also present limitations, such as generating inaccurate or ’hallucinated’ information capable of negatively influencing decision-making processes. This thesis investigates the behavior of Large Language Models in cybersecurity risk assessment scenarios compared to human experts, highlighting the importance of human oversight and the integration of humans and machines in this process. To achieve this, a questionnaire containing risk scenarios was applied to human participants, whose responses were compared with those generated by five popular Large Language Models. The results suggest that LLMs tended to underestimate security risks compared to human participants. Therefore, although they may be useful as support tools, their use as autonomous evaluators in information security risk assessment processes should be considered with caution. 2026-07-08T00:00:00Z https://repositorio.ufu.br/handle/123456789/49999 Title: Um algoritmo genético transgênico para otimização de hiperparâmetros de redes long short-term memory aplicado à previsão de preços de café arábica Abstract: Financial time series forecasting remains a challenging problem due to the high vo- latility, nonlinearity, and complexity of financial markets. This dissertation investigates the use of Long Short-Term Memory (LSTM) neural networks to forecast Arabica coffee futures closing prices, aiming to analyze the contribution of evolutionary hyperparameter optimization to the development of models capable of balancing predictive performance, financial performance, and computational cost. Three approaches were compared: a ba- seline LSTM with fixed hyperparameters, an LSTM optimized by a Genetic Algorithm (LSTM-GA), and an LSTM optimized by an adapted Transgenetic Genetic Algorithm (LSTM-TGA). The proposed methodology incorporates a probabilistic genetic history and four transgenetic modules into the evolutionary process in order to modify the se- arch dynamics for promising hyperparameter configurations. Experiments were conducted using historical closing prices of Arabica coffee futures contracts (KC=F) considering th- ree temporal resolutions: 4-hour, daily, and weekly. Model evaluation included predictive accuracy metrics, convergence and population diversity analyses, architectural complexity, computational cost, architectural efficiency, and financial validation through backtesting. The results demonstrate that evolutionary optimization effectively automated hyperpa- rameter selection, producing competitive neural architectures and showing that different temporal resolutions require distinct configurations. Furthermore, improvements in pre- dictive performance were not necessarily associated with superior financial performance, highlighting the importance of an integrated evaluation involving predictive accuracy, financial performance, architectural complexity, and computational cost. The adapted TGA proved to be a competitive optimization strategy by generating structurally simpler architectures, preserving population diversity, and producing models with high architec- tural efficiency. It is concluded that the proposed methodology represents a promising approach for automatic LSTM hyperparameter optimization, contributing to the advan- cement of evolutionary optimization applied to Artificial Intelligence and financial time series forecasting. 2026-07-30T00:00:00Z