A Comparative Analysis of SHAP, LIME, ANCHORS, and DICE for Interpreting a Dense Neural Network in Credit Card Fraud Detection

Financial institutions heavily rely on advanced Machine Learning algorithms to screen transactions. However, they face increasing pressure from regulators and the public to ensure AI accountability and transparency, particularly in credit card fraud detection. While ML technology has effectively detected fraudulent activity, the opacity of Artificial Neural Networks (ANN) can make it challenging to explain decisions. This has prompted a recent push for more explainable fraud prevention tools. Although vendors claim to improve detection rates, integrating explanation data is still early. Data scientists recognize the potential of Explainable AI (XAI) techniques in fraud prevention, but comparative research on their effectiveness is lacking. This paper aims to advance the comparative research on credit card fraud detection by statistically evaluating established XAI methods. The goal is to explain and validate the fraud detection black-box machine learning model, where the baseline model used for explanation is an ANN trained with a large dataset of 25,128 instances. Four explainability methods (SHAP, LIME, ANCHORS, and DiCE) are utilized, and the same test set is used to generate an explanation across all four methods. Analysis through the Friedman test indicates a statistical significance of the SHAP, ANCHORS, and DiCE results, validated with interpretability and reliability aspects of explanations such as identity, stability, separability, similarity, and computational complexity. The results indicated that SHAP, LIME, and ANCHORS methods exhibit better model interpretability regarding stability, separability, and similarity.