Fraud detection using TabNet* classifier: a machine learning approach

Abstract

Detecting fraudulent transactions is a big challenge in the digital financial world. Transaction volumes are growing quickly, and new attack methods often outstrip traditional detection systems. Current fraud-detection models usually lack clarity and do not perform reliably on unbalanced real-world datasets. This highlights the urgent need for clear and explainable deep-learning methods for tabular financial data. This paper presents an interpretable deep learning framework built on the TabNet classifier. It uses attention-driven feature selection, sparse representation learning, and sequential decision reasoning to model complex interactions among transactional, demographic, and geographical factors. The model was tested on a real-world credit card transaction dataset with 23 features. It achieved 99.69% accuracy, a 0.975 F1-score, and a 0.956 ROC-AUC. This performance outperforms benchmark models such as random forest, XGBoost, LightGBM, and logistic regression. In addition to outstanding predictive results. Furthermore, interpretability is enhanced by TabNet's attention-based feature attribution. This facilitates the clear understanding of model decisions, supporting its use in regulated financial environments where precision and responsibility are crucial.