https://repositorio.ufu.br/handle/123456789/17904 2026-09-19T00:48:01Z https://repositorio.ufu.br/handle/123456789/50132 Title: Análise comportamental com aprendizado de máquina para detecção de atividades suspeitas em sistemas corporativos Abstract: This work proposes and implements a behavioral analysis system using Machine Learning for detecting suspicious activities in a coffee brokerage management software. A supervised classification model built with TensorFlow.js is compared against a static rules baseline. The system combines a Node.js/Express/SQLite backend for model training with a React/TypeScript frontend for real-time browser-side inference. A dense neural network with 4,993 trainable parameters was trained on 10,000 synthetic samples across eight behavioral profiles using 30 numerical features. A Human-in-the-Loop mechanism allows administrators to review incidents and feed incremental retraining. The model achieved an F1-Score of 99.68% against 82.16% for the baseline, with recall of 99.48% versus 71.98%, producing only 4 false negatives in 1,500 test samples compared to 216 for the baseline. Results reject the null hypothesis of equivalence between the two approaches, with the caveat that data are synthetic and the baseline was defined within the work itself. Keywords: Machine Learning, Behavioral Analysis, Supervised Classification, Information Systems Security, Anomaly Detection. Notes: Código-fonte do sistema disponível em: https://github.com/KAYOKG/PROJETO-TCC-UFU 2026-07-31T00:00:00Z https://repositorio.ufu.br/handle/123456789/50126 Title: Análise do posicionamento político dos brasileiros em relação à guerra Irã x EUA utilizando dados do YouTube Abstract: The growing volume of data produced on social networks has expanded the possibilities for investigating sociopolitical phenomena and public opinion behavior. This study aims to analyze the positioning of Brazilians regarding the war between the United States and Iran using comments extracted from YouTube. The methodology involved collecting 14,052 comments from videos related to the topic, followed by automated labeling of the texts regarding their stance (pro, against, or neutral to the war) and the supported side (US or Iran) using the Gemini 3.1 Pro large language model. For the structural analysis of the discussions, the data was modeled as undirected and weighted complex networks, whose edges represent lexical similarity between comments, projected from a bipartite comment-term structure. The application of centrality metrics (Degree, Betweenness, Closeness, and PageRank) identified a small core of comments that exert strong structural influence on the discussions. Additionally, the Girvan-Newman algorithm detected the formation of 10 distinct communities, revealing blocks of discourse with shared themes and vocabularies. The results demonstrate the effectiveness of combining text mining techniques and complex network theory to map and interpret spontaneous user manifestations, showing that relational structures can assist in identifying opinion patterns in digital political debate. 2026-08-26T00:00:00Z https://repositorio.ufu.br/handle/123456789/50012 Title: SQL vs. NoSQL: um estudo comparativo de desempenho em operações CRUD com PostgreSQL e MongoDB 2026-07-13T00:00:00Z https://repositorio.ufu.br/handle/123456789/49740 Title: Diagnóstico de Alzheimer baseado em imagens de ressonância magnética utilizando redes convolucionais 2025-05-05T00:00:00Z