Da Teoria à Prática: Metodologias de Treinamento para Prototipagem Eficiente de Modelos de Árvore de Decisão

This study explores the application of decision trees in classification tasks, using the Iris and BreastCancer datasets. Hyperparameter configuration with GridSearch, different validation strategies (holdout, K-Fold, and stratified cross-validation), and the impact of the SMOTE technique in addressing class imbalance, specifically in the Breast Cancer dataset, were investigated. Adopting the rigorous CRISP-DM methodology, the work emphasized exploratory data analysis for identifying relevant attributes, clear prototyping, and data leakage prevention, aiming to ensure the robustness of supervised learning models. The results indicate that hyperparameter optimization, such as maximum depth and minimum samples per leaf, along with SMOTE, allowed for the construction of robust and generalizable decision tree models (based on the CART algorithm). The use of SMOTE was particularly effective in increasing sensitivity in the classification of malignant tumors, significantly reducing type II errors, a critical aspect in the clinical context. The interpretability of the models and the identification of highly relevant predictive features, corroborated by statistical and visual analyses, were highlighted aspects. The entire experiment was developed in Google COLAB, ensuring the reproducibility of the results.