TEMPORAL GRAPH NEURAL NETWORKS FOR SEQUENTIAL ANOMALY DETECTION IN REAL-TIME E-COMMERCE STREAMS

The exponential growth of e-commerce transactions has created an urgent need for sophisticated anomaly detection systems capable of identifying fraudulent activities, system malfunctions, and unusual behavioral patterns in real-time data streams. Traditional anomaly detection approaches fail to capture the complex interdependencies between entities and the temporal evolution of their relationships within e-commerce ecosystems. This paper presents a novel framework that integrates Temporal Graph Neural Networks (TGNNs) with advanced graph representation learning techniques to address sequential anomaly detection in real-time e-commerce environments. Our approach leverages the structural modeling capabilities of Graph Neural Networks (GNNs) while incorporating temporal dynamics through specialized attention mechanisms and incremental learning strategies. The framework employs a multi-scale graph construction process that captures both local neighborhood structures and global network patterns, enabling the identification of anomalous nodes and subgraphs that deviate from established community structures. We introduce an adaptive random walk strategy inspired by Node2Vec that balances breadth-first and depth-first exploration to capture diverse types of anomalous patterns across different temporal scales. Comprehensive evaluation on three large-scale e-commerce datasets demonstrates significant performance improvements, with our method achieving 17.2% enhancement in F1-score and 14.6% improvement in Area Under Curve (AUC) compared to state-of-the-art approaches, while maintaining sub-second inference times suitable for real-time deployment.