Increasing Transaction Fraud Prediction Ability by Using Multi-Task Learning and Pruning

Abstract

Transaction fraud detection is an important subject especially for financial institutions. If fraudulent transactions are not prevented, results can be harmful. Consequences are not only financial but also loss of brand reputation and customer loyalty. Rule based methods are widely used to prevent fraudulent transactions, these methods require in-depth business knowledge and past experience, usually rules are generated after fraud events occur. As machine learning methods are an alternative in this problem, research are carried out to overcome the difficulties that arise with unbalanced class distribution. It is seen that some techniques such as data augmentation, synthetic data generation or reduction of data in the majority class are used before the use of discriminant models. Generative models such as auto-encoders, which are widely used in anomaly detection, can also offer a solution for this problem. In this research, an auto-encoder and a classifier which is connected to the bottleneck layer are trained simultaneously in the same model in supervised manner. It is aimed to increase the predictive ability of the model by applying pruning techniques and comparable results are obtained and shared.