An Optimized Machine Learning Framework for Financial Fraud Detection Using SMOTE and Feature Selection

Abstract

This study presents a machine learning system capable of efficiently detecting financial fraud with the use of unbalanced data treatment strategies. A holistic approach to the problem, combining feature scaling, feature selection, SMOTE (Synthetic Minority Oversampling Technique), and machine learning evaluation was proposed by all in a single experimental pipeline. Method119 randomly over-samples 84 million transactions that occur on average every day with only 2 out of 1000 transactions reflecting fraud241. Model Evaluation- The model accuracy, precision, recall and F1-score and ROC-AUC. In the results of their test, they demonstrate Random Forest won out among all numeric methods with 99.96% accuracy, 97% precision and 91% recall (Contextual F1 = 94 & ROC-AUC =98). The results indicate the effectiveness of combining SMOTE with feature selection to achieve better detection for frauds and balance the classification towards lower outputs, i.e., which are Genuine Transactions. We provide new insights into the reliability of machine learning-based fraud detection system in real financial systems and a general framework for implementing these models.