Explainable Personalized Behavioural Fraud Detection Using Deviation-Based Feature Engineering and Deep Learning

Financial fraud continues to be a major problem in digital payment systems.  This work introduces a personalized system for detecting fraud that uses behaviour patterns, builds features by looking at how they differ from usual activity, and applies a multilayer perceptron along with SHapley Additive exPlanations to help explain the results clearly.  Using the PaySim dataset, user profiles and normalized deviation features help detect unusual behaviour patterns.  The model has a high accuracy of 99.5%, good precision of 56.4%, strong recall of 95.8%, a fair F1-score of 71.5%, and a very high area under the curve of 99.8%, showing it can detect fraud accurately and clearly explain its findings.