TcfDyG: Cryptocurrency Network Anomaly Detection Based on Time-aware Channel Fusion Dynamic Graph Neural Network

With the rapid advancement of blockchain technology and cryptocurrency systems, transaction behaviors in financial networks have become highly dynamic and exhibit complex evolutionary patterns. However, effectively modeling these behaviors faces a significant challenge: transaction data contains heterogeneous information that evolves intricately over time. Existing methods often struggle to capture the deep, nonlinear correlations among these multi-channel features, limiting their ability to uncover latent risks. Therefore, how to unify temporal evolution with channel-wise feature interaction to enable accurate risk detection has become a focal point in current research. In this paper, we propose a T ime-aware C hannel F usion Dy namic G raph neural network ( TcfDyG ) for modeling dynamic evolution and performing link prediction in cryptocurrency networks. Specifically, we employ a time-aware historical neighbor sampling mechanism to extract multimodal features, including node attributes, transaction edge features, temporal encodings, and co-occurrence-based neighbor interaction information. Subsequently, we integrate recurrent neural networks for temporal dependency modeling with a multilayer perceptron-based hybrid module to perform channel-wise feature interaction. Extensive experiments conducted on eight real-world datasets demonstrate that TcfDyG significantly outperforms state-of-the-art baselines in link prediction tasks. The results demonstrate that the proposed integrated framework effectively improves dynamic representation learning by jointly modeling temporal evolution and heterogeneous feature interactions, while maintaining a favorable trade-off between predictive accuracy and computational efficiency. Our code is available at https://github.com/ly896/TcfDyG .