Um estudo sobre diferentes métodos de classificação para dados extremamente desbalanceados sob a ótica do aprendizado sensível ao custo

Abstract

In classification tasks, we often encounter datasets with highly imbalanced classes, such as those found in rare disease diagnosis, financial fraud detection, and industrial failure analysis. Training traditional models on such datasets tends to favor the majority class, resulting in poor ability to identify observations from the minority class, which is usually the one of greatest interest. In this undergraduate thesis, we present and compare different classification models — Logistic Regression, Random Forest, and XGBoost — from the perspective of cost‑sensitive learning, applying class weights to compensate for severe imbalance. Using a real‑world credit card transaction dataset (with only 0.172% fraud cases), we demonstrate how assigning differential weights to classification errors significantly improves the sensitivity (recall) in detecting the minority class. Furthermore, we evaluate which combination of model and validation technique (data‑split or stratified cross‑validation) yields the best trade‑off between sensitivity, precision, and F1‑score. The results indicate that XGBoost with appropriate weighting achieves the best overall performance, reaching recall above 0.89 and F1‑score close to 0.88.