Deep Learning vs. traditional Machine Learning algorithms used in Credit Card Fraud Detection

With the continuing growth of E-commerce, credit card fraud has evolved exponentially, where people are using more on-line services to conduct their daily transactions. Fraudsters masquerade normal behaviour of customers to achieve unlawful gains. Fraud patterns are changing rapidly where fraud detection needs to be re-evaluated from a reactive to a proactive approach. In recent years Deep Learning has gained lot of popularity in image recognition, speech recognition and natural language processing. This paper seeks to understand how Deep Learning can be helpful in finding fraud in credit card transactions and compare Deep Learning against several state of the art algorithms (RF, GBM, GLM) and sampling methods (Over, Under, Hybrid, SMOTE and ROSE) used in fraud detection. The results show that Deep Learning performed best with the highest Recall (accuracy of identifying fraudulent transactions), which means lowest financial losses to the company. However, Deep Learning achieved the lowest Precision rate (classified more legitimate transactions as fraudulent), which can cause customer dissatisfaction. Among other chosen classifiers, oversampling method performed best in terms of AUC, precision was highest for GLM and F-Score was highest for model trained using ROSE sampling method. Recall and Precision both have high cost, so there cannot be any trade of one against the other. Selecting the best classifier to identify fraud is based on the business goal. Keywords—Imbalanced class, Data mining, Sampling methods, H2O, Gradient Boosting, Random Forest, Generalized Linear Models, Deep Learning, Grid search, Hyper parameter optimization, Ensemble methods.