Trustworthiness Validations of Machine Learning Models for Fraud Detection in Mobile Based Business Transactions Using XAI

Abstract

The increasing usage of mobile-based transactions makes the business operations more in need of strong fraud detection systems. However, due to the black-box nature of several of the advanced machine learning models, trust in such automated fraud detection has been limited because stakeholders lack insight into how the model provided the decisions. This paper discusses the trustworthiness validation of Machine Learning (ML) models such as Random Forest (RF) and MLP Classifier (Multilayer Perceptron Model) and XGBoost (Extreme Gradient Boosting) applied for the detection of fraudulent transactions within the mobile-based business transactions by using Explainable Artificial Intelligence (XAI) techniques. Some of the most widely used XAI models applied in this work are LIME, SHAP, Anchor Explanations, Surrogate Models, and Explainable Boosting Machine (EBM). These XAI models assess the trustworthiness of the ML models using the most crucial metric such as fidelity and the applied XAI models are validated using unambiguity and interpretability size. The implication of the findings is a framework through which the reliability of models enhancement is done in order to work with increased confidence in ML-driven fraud detection systems within the mobile payment sector.