Enhancing Fraud Detection in Digital Transactions through a Modified CART Algorithm

Abstract

Online payment fraud has become a major challenge in the banking and e-commerce sectors due to the rapid growth of digital transactions and increasingly sophisticated cybercrime techniques. Although the Classification and Regression Trees (CART) algorithm has been widely applied to fraud detection, its performance may deteriorate when dealing with imbalanced datasets. This study proposes CART-SQ (Classification and Regression Trees Simple Quadrat), which applies an equivalent quadratic formulation of the Gini impurity during attribute split evaluation while preserving the mathematical properties of the conventional CART algorithm. The proposed method was evaluated using an imbalanced financial transaction dataset and compared with the conventional CART and C4.5 algorithms under two train–test partition scenarios. The experimental results demonstrate that CART-SQ achieved the best overall classification performance, with accuracies of 98.25% and 98.20%, while also attaining the highest precision and F1-score for the fraud class. These findings indicate that the proposed equivalent quadratic formulation provides an effective alternative for improving decision tree performance in fraud detection tasks involving imbalanced datasets.