XFraudNet: Explainable and Adaptive Deep Temporal Network for Transaction Anomaly Detection

• Shabiya M I,
• R Roopa Chandrika,

Vol. 19, No. 12, December 31, 2025
10.3837/tiis.2025.12.010, Download Paper (Free):

• Credit Card Fraud Detection
• Explainable Deep Learning
• Temporal Feature Modeling
• Bayesian Hyperparameter Optimization
• Hybrid Neural Network Architecture

Abstract

The rapid expansion of digital financial services has led to a surge in credit card fraud, necessitating robust, real-time, and interpretable systems for fraud detection. An outline of the XFraudNet prototype, a hybrid deep learning construct, begins with the 1D convolutional neural networks for local pattern recognition, and proceeds to the Bidirectional long short-term memory units for temporal sequencing and an attention mechanism describing the retrieval of pertinent features within a timeline. XFraudNet seeks to close the explainability gap, targeting deep models’ interpretability by SHAP Explainable Artificial Intelligence, and thus, allowing for explainable deep learning within the confines of the law. Employing Bayesian Optimization, hyperparameter fine-tuning is achieved to ensure model generalization on the datasets, which are imbalanced. XFraudNet’s experimental evaluation on multiple datasets reveals an F1-score (99.0%), AUC (99.17%), precision (98.68%), and recall (99.33%), which signifies an unsurpassed performance, thus demonstrating the model’s superiority to other classical machine learning models and contemporary ones like TabNet, detectGNN, and GAT-COBO. Each component’s importance within the architecture has been validated through extensive ablation studies. The model’s explainability, coupled with its lightweight, emphasizes its importance and real-time usage. With the fine-tuning of explainability and other constructs, the model offers a practical and scalable approach to fraud detection.

Statistics

Cite this article