An Adaptive Temporal Convolutional Network Framework for Real-Time Credit Card Fraud Detection in Highly Imbalanced Financial Transaction Data

Abstract

This paper focuses on credit card fraud detection using Temporal Convolutional Networks (TCN), real-time fraud detection, imbalanced data handling, and comparative performance against Logistic Regression, Random Forest, and XGBoost. The rapid expansion of digital payments has increased the need for fraud detection systems that can operate with low latency, maintain high predictive performance, and provide interpretable decisions for financial analysts. This study develops an adaptive TCN-based framework for real-time credit card fraud detection using a highly imbalanced transaction dataset. The framework integrates preprocessing, feature engineering, chronological validation, imbalance-aware learning, comparative baseline modelling, low-latency inference testing, and SHAP-based explainability. The experiment used the Kaggle European credit card fraud dataset containing 284,807 transactions, of which 492 were fraudulent, representing a fraud rate of 0.173 percent and an imbalance ratio of 577.88:1. Results show that Random Forest achieved the highest PR-AUC of 0.773 and precision of 0.925, XGBoost achieved the highest ROC-AUC of 0.977 with p99 latency of 6.40 ms, while the TCN achieved competitive fraud recall of 0.808, PR-AUC of 0.714, ROC-AUC of 0.974, and p99 latency of 6.34 ms. SHAP explanations identified stable fraud-predictive patterns across models, supporting transparent fraud investigation. The findings demonstrate that adaptive, explainable, and low-latency fraud detection frameworks can strengthen financial transaction monitoring in modern digital payment environments.