Comparing ROC Curve based Thresholding Methods in Online Transactions Fraud Detection System using Deep Learning

Data imbalance is one of the main issues in online transactions fraud detection systems where the majority class (i.e., genuine transactions) dominates the minority class (i.e., fraud transactions). Deep learning-based fraud detection systems have the capability to learn efficiently with imbalanced data. Thresholding with Deep Learning is still understudied. Thresholding or threshold moving alters the decision threshold of the learning model to change its output rather than changing the data distribution or learning of the model. A fraud detection system is considered efficient if it can detect the maximum number of frauds and maintain the overall performance of model. Thus, this paper presents a comparison of three thresholding methods based on Receiver Operating Characteristic (ROC) Curve i.e., Closest to (0,1) criteria, Youden Index (J), max -G-Mean in deep-learning based online transactions fraud detection system. The experimental results show that closest to (0,1) criterion has achieved maximum fraud detection rate i.e. True Positive Rate (TPR) as compared to the other two methods. Hence, selecting the right decisionthresholding method helps in achieving better results.