Advanced Machine Learning Models for Credit Card Fraud Detection

Abstract

Credit card fraud detection is a cost-sensitive learning problem characterized by extreme class imbal

ance, non-stationary adversarial behavior, and stringent operational constraints on false alarms. Using

the publicly available CREDITCARDFRAUD-ULB benchmark of European cardholder transactions, we

develop and evaluate a family of advanced deep learning models specialized for continuous tabular data.

Our framework combines three ingredients: (i) a cost-sensitive residual multilayer perceptron (RESMLP)

that provides a strong supervised baseline; (ii) a feature-tokenizing transformer (FT-TRANSFORMER) that

contextualizes each transaction attribute through self-attention; and (iii) an innovative self-supervised pre

training strategy (FRAUDCL-FTT) that couples masked feature modeling with contrastive representation

learning prior to supervised fine-tuning. We formulate fraud detection as a calibrated risk scoring problem

and therefore evaluate models not only by ranking metrics such as AUROC and AUPRC, but also by

probability calibration and decision-theoretic threshold selection. The resulting manuscript is designed to

be fully reproducible: the accompanying code automatically trains the models, computes metrics, and

exports publication-ready tables and figures.