Innovative novel regularized memory graph attention capsule network for financial fraud detection

Abstract

Financial fraud detection (FFD) is crucial for ensuring the safety and efficiency of financial transactions. This article presents the Regularised Memory Graph Attention Capsule Network (RMGACNet), an original architecture aiming at improving fraud detection using Bidirectional Long Short-Term Memory (BiLSTM) networks combined with advanced feature extraction and classification algorithms. The model is tested on two reliable datasets: the European Cardholder (ECH) transactions dataset, which contains 284,807 transactions and 492 fraud instances, and the IEEE-CIS dataset, which has more than 1 million transactions. Our approach enhances comparison to existing methods of feature selection and classification accuracy. On the ECH dataset, RMGACNet achieves an accuracy of 0.9772, a precision of 0.9768, and an F1 score of 0.9770 measures; on the IEEE-CIS dataset, it achieves an accuracy of 0.9882, a precision of 0.9876 and an F1 score of 0.9879. The findings indicate that RMGACNet routinely surpasses existing models’ efficiency and accuracy while ensuring strong execution time performance, especially when handling large-scale datasets. The suggested model demonstrates scalability and stability, making it suitable for real-time financial systems.