Uncertainty Aware Based Deep Learning for Credit Card Fraud Detection

Inhalt:
By designing different network topologies or learning models, many research efforts on deep neural networks have focused on improving the accuracy of endpoint predictions and reducing undesired biases in the task of credit card fraud detection. Quantifying uncertainty, as well as point estimates, is important because it lowers model unfairness and aids practitioners in developing dependable systems that prevent making bad decisions due to a lack of confidence. In realworld credit card fraud detection settings, evaluating the uncertainty associated with DNN predictions is critical because fraudsters constantly change their techniques, and as a result, DNN encounters observations that differ from the training distribution, and due to the time-consuming process, few transactions are checked by professional experts in time to update the DNN. As a consequence, this work provides Monte Carlo dropout, ensemble, and ensemble Monte Carlo dropout as three uncertainty quantification methodologies for card fraud detection on transaction data. Forecast uncertainty estimates were also assessed using the UQ confusion matrix and a variety of performance criteria. We show that the ensemble captures the uncertainty associated with generated predictions experimental measurements better than the individual. We also demonstrate that the proposed UQ approach enhances fraud detection by giving more information for point prediction.