Authors: N. Rajarajeswari, Dr. Saley Seetharaman
Abstract: The explosive increase in the use of digital payment systems has led to an increase in complex financial scams. It is necessary to have intelligent detection and risk assessment systems that can detect and prevent potential financial crimes in real time. However, deep learning models have proven to be accurate, but their lack of transparency makes them unsuitable for highly regulated financial systems. In this study, we present a novel Explainable AI system to address fraud detection and risk assessment in real time. Our system includes a Temporal Fusion Transformer network model for sequence-aware fraud detection and a hybrid explainability module that incorporates feature-based SHAP, locally-explainable LIME, and counterfactually-generative rules to provide actionable explanations. We have evaluated our system on a real-world dataset comprising 10 million digital payments. Our experiments show that the proposed TFT achieves an area under curve (AUC) of 0.982, significantly better than XGBoost (0.965) and LSTM (0.971).
DOI: https://doi.org/10.5281/zenodo.20411601
