Adaptive Spatio-Temporal Aggregation for Temporal Dynamic Graph-Based Fraud Risk Detection

Abstract

This paper introduces an advanced fraud detection algorithm, AT-GCN, tailored for temporal dynamic graphs frequently encountered in financial domains such as money laundering and financial fraud. Traditional graph neural networks (GNNs) have been predominantly successful in static graph analysis, lacking the capability to capture the temporal dynamics of transactions. To address this, the proposed AT-GCN algorithm integrates three key innovations: adaptive parameter updates using LSTM to reflect the temporal evolution of graph structures, a resampling method across time steps to balance label distribution and leverage temporal correlations, and a novel similarity-based weighted aggregation approach that enhances the differentiation of node importance within the graph. The LSTM component allows the model to dynamically adjust to changes in graph topology, capturing temporal dynamics effectively. The oversampling strategy mitigates label imbalance by connecting nodes across various time steps using Euclidean distance, enriching the model's fraud detection accuracy. The aggregation method, underpinned by a machine learning perceptive model, assigns weights based on node similarity, thereby tackling the disguise problem posed by fraudulent entities. Empirical results demonstrate AT-GCN's superiority over existing methods, showcasing its potential in real-world dynamic graph applications.