An enhancement of classification and regression tree (CART) algorithm in the implementation of SMS fraud detection

ABSTRACT: The increasing prevalence of fraudulent SMS necessitates robust detection systems capable of accurately fraud instances. While the traditional CART algorithm offers interpretability and simplicity, it struggles with challenges such as overfitting, high variance, and class imbalance, often favoring the majority class. This study proposes an enhanced CART framework by integrating Mutual Information (MI) for feature selection, bagging for variance reduction, and SMOTE for handling class imbalance. The MI-enhanced CART model demonstrated improved classification reliability, increasing accuracy from 94% to 95%, with corresponding improvements in precision (78% to 80%), recall (78% to 80%), and F1-score (0.78 to 0.80). Incorporating bagging further boosted performance, raising precision from 81% to 87%, recall from 75% to 78%, and F1-score from 0.77 to 0.82. Finally, the integration of SMOTE effectively addressed class imbalance, raising the F1-score for spam from 0.78 to 0.82 and maintaining strong recall at 77%. Evaluation using metrics such as precision, recall, F1-score, and AUC-PR confirms the effectiveness of this composite approach. These enhancements demonstrate the combining MI, bagging, and SMOTE within the CART model significantly improves the detection of fraudulent SMS, offering a solution for fraud detection.

Filipiniana