Development and Сomparison of Machine Learning Approaches for Fraud Detection in Banking Institutions

This thesis addresses the automatic detection of transaction frauds in banking institutions by means of machine learning application. An evaluation of the acknowledged supervised techniques is conducted to demonstrate the preferable algorithm using general classification metrics as well as receiver operating characteristics using a synthetically created set of information. The comparison is completed by the inclusion of an individually developed methodology focusing on the derivation of rules through an analysis of similar feature values. It is shown that the gradient boosted tree algorithm outperforms other examined methods in terms of accuracy, recall, F-score, Cohen’s kappa and area under curve value. Although the individual approach provides inferior outcome to the preferred acknowledged technique, it enables advantages in terms of controllability of mistakes and losses as well as comprehensibility, which represent essential characteristics for the concept selection of banks because of their need to balance benefits and risks in an optimal way. Due to higher influenceability compared to algorithms operating as black boxes and its generally good results, the customized variant constitutes an excellent alternative for payment service providers.