CREDSHAP- Credit Card Fraud Detection using SHAP

The online banking and digital payments industry is experiencing significant growth that has unavoidably accompanied a rise in credit card fraud. This is a problem global in nature and concern. Nowadays, the detection of fraud patterns through traditional rule-based systems is often quite rigid and hence, very difficult, or even impossible, to spot. Furthermore, even if deep learning models were to be used for fraud detection, they would still suffer from lack of transparency and that could be a big drawback in certain regulated sectors. Therefore, the present study proposes the new CredSHAP system, which is an interpretable deep learning model. It combines a feed-forward neural network with Shapley Additive exPlanations (SHAP). The model has been trained on and evaluated with an imbalanced dataset of credit card transactions that is popular among researchers, where the proportion of fraudulent transactions is less than 0.2%. Random oversampling was applied to deal with the imbalance in the dataset while the Adam optimizer was used for the model optimization with binary crossentropy loss. CredSHAP produced results of 99.3% accuracy, 97.8% recall, and an F1-score of 0.96, which are all superior to the performance of the logistic regression and random forest baselines. Moreover, the use of SHAP adds value to the model by shedding light on the contribution of each feature to the prediction, thus giving the analysts the required transparency. The combination of accuracy and explainability thus creates a situation where trust in AI-based fraud detection systems is highly strengthened. The researchers plan to add ensemble learning and real-time processing to the system for large-scale proposals in future work.