Towards Transparent Fraud Detection: Explainable AI and Multi-algorithm Optimization in Financial Security

Abstract

As a result of these payment methods being used so often, the incidence of credit card fraud has grown. Electronic payment systems have improved and e-commerce has grown in popularity as a result of credit card usage. A more extensive use of machine learning procedures is developed for the purpose of detecting and preventing fraud. When it comes to evaluating consumer data, ML algorithms are crucial. In the context of credit card fraud detection using deep learning models, the properties of explainable artificial intelligence (XAI) offer versatile and multidimensional possibilities. The goals of this study were to identify the decision-making process of sophisticated artificial intelligence (AI) models and to interpret and explain the decisions made by these models. The input data was cleaned, tokenized, and lemmatized as part of the data pre-treatment to remove any inconsistencies. In an effort to streamlined. In order to find new insights and patterns in the datasets, exploratory data analysis was carried out. To classify the data and determine if fraud has taken place, the research employs Explainable AI models like Local Interpretable Model-Agnostic Explanations (LIME) and Shapley additive explanation (SHAP), in addition to transformer-based models like BERT, XL Net, and RoBERTa. The study presents CheMO, a method of multi-algorithm optimization that draws inspiration from the game of chess, to help with parameter tuning. In CheMO, different pieces stand for different optimization methods. Pawns represent Simulated Annealing, Bishops for the BFGS algorithm, Knights for the Nelder-Mead (NM) method, Rooks for the OAT optimization method, and Queens for a multi-algorithm optimization that follows the sequence BFGS-NM-OAT. The King stands for the algorithm's top performance. Afterwards, the model's performance was assessed with the use of F-measure metrics, Precision, Accuracy, and Recall. The big dataset (284, 807 characteristics and 492 fraudulent transactions) was a major factor in the high accuracy rates (97 to 98%) shown by all models. However, accuracy alone was insufficient to offer a thorough comparison metric. Finally, in order to understand the model's decision-making procedure better, explainable AI modeling was used.