Behavioral--Transactional Data Analytics using Hierarchical Feature Fusion with Attention Deep Learning Model for E-Commerce Fraud Detection

Abstract

The rapid growth of e-commerce has intensified fraudulent activities, posing a serious threat to online platforms by distorting ranking systems and degrading user engagement, making fraud detection a task of high real-world importance despite the adversarial responses of fraudsters. Existing fraud detection systems in e-commerce are largely static, relying on previously identified fraud behaviors as supervision data and struggling to adapt to evolving fraud patterns. Although developments in artificial intelligence, machine learning, and cloud computing have strengthened this field, existing ML and data mining approaches applied to platforms such as Facebook and eBay remain insufficiently complex to capture evolving fraud behavior. This study presents an E-Commerce Fraud Detection Framework using a Multi-Graph Neural Network with Attention (ECFD-MGNNA), designed to accurately classify fraudulent and legitimate transactions. The proposed method first pre-processes the input data through outlier detection, missing-value handling, and feature scaling to standardize inputs and improve training efficiency. A fusion feature selection strategy is then applied, integrating Pearson correlation analysis, variance filtering, and mutual information to retain highly relevant, informative features while eliminating redundancy. The selected features are subsequently passed into a multi-graph neural network attention fusion model for robust transaction classification. The ECFD-MGNNA system is experimentally evaluated on two benchmark fraud detection datasets using standard performance metrics, and the simulation results demonstrate its superior performance over recent state-of-the-art methodologies.