This study focuses on the development of machine learning models for detecting fraud in financial transactions, with particular attention to addressing key challenges such as class imbalance, high false positive rates, and the need for real-time detection. A novel fraud detection model referred to as RFS—an integration of Random Forest, SMOTE (Synthetic Minority Oversampling Technique), and feature selection methods—is proposed to enhance both accuracy and interpretability. Feature selection was carried out using Recursive Feature Elimination (RFE), Mutual Information (MI), and SHAP (Shapley Additive Explanations) to identify and retain the most informative variables, including transaction amount, frequency, and previous fraud history. The Random Forest classifier trained on the optimized feature set and balanced with SMOTE, achieved notable performance with 99.84% accuracy and a 0.979 ROC-AUC score. Feature selection not only reduced over fitting risks but also improved recall for fraudulent classes, contributing to a more effective detection model. The results demonstrate that combining feature engineering, data balancing, and an interpretable learning algorithm significantly improves fraud detection capabilities. The proposed RFS model offers a practical and scalable solution for financial institutions seeking to enhance digital transaction security and reduce economic losses due to fraud.
© 2026 IJARIE. All Rights Reserved.
Designed by LexyCrest Solutions