SHAP-Enhanced Hybrid Model for Accurate Financial Fraud Detection

The detection of fraudulent transactions presents challenges due to the nature of fraudulent datasets, where fraud cases hardly exist against legitimate ones. In the author's opinion, other traditional approaches performed significantly better but failed to meet the high quality and accuracy required for practical deployment. To face these issues, this work proposes a hybrid machine learning framework that combines Isolation Forest for anomaly detection and Logistic Regression for supervised classification purposes. The class imbalance issue is handled by SMOTE, which creates fake samples of the minority group. Besides that, some kind of interpretability is provided to the model through the application of SHAP (SHapley Additive exPlanations), which gives an indication of which features are influential in contributing towards the fraud predictions. The framework shows high precision, with a much lower rate of false positives and provides transparency that makes it possible to scale and optimally solve the gap in practical applicability for fraud detection in real-world scenarios.