Uncertainty-aware credit card fraud detection using deep learning

Countless research works of deep neural networks (DNNs) in the task credit card fraud detection have focused on improving accuracy point predictions and mitigating unwanted biases by building different network architectures or learning models. Quantifying uncertainty accompanied estimation is essential because it mitigates model unfairness permits practitioners to develop trustworthy systems which abstain from suboptimal decisions due low confidence. Explicitly, assessing uncertainties associated with DNNs critical real-world settings for characteristic reasons, including (a) fraudsters constantly change their strategies, accordingly, encounter observations that are not generated same process as training distribution, (b) owing time-consuming process, very few transactions timely checked professional experts update DNNs. Therefore, this study proposes three quantification (UQ) techniques named Monte Carlo dropout, ensemble, ensemble dropout applied transaction data. Moreover, evaluate predictive estimates, UQ confusion matrix several performance metrics utilized. Through experimental results, we show more effective capturing corresponding predictions. Additionally, demonstrate proposed methods provide extra insight predictions, leading elevate prevention process.