DyLF: Introducing Longformer for More Efficient Dynamic Graph Learning

In scenarios with fixed time intervals, discrete-time dynamic graph (DTDG) is used to represent dynamic networks, such as social networks, traffic networks and biological networks. Existing models introduce self-attention mechanism for structural representation learning on discrete-time dynamic graph and demonstrate good performance in this field. However, temporal encoder in these models focuses on the entire time domain, introducing noise and limiting the model's ability to extract critical information. To address this issue, we propose the DyLF model based on Longformer mechanism. By incorporating the Longformer-based self-attention mechanism, DyLF can flexibly adjust the range of visible time domain while encoding temporal features of nodes in dynamic graphs and efficiently capture temporal dependencies between snapshots, thereby enhancing the ability to extract more effective temporal information. Extensive experiments show that DyLF achieves an average improvement of 4.28% in AUC and 6.87% in F1_score with the existing baseline, demonstrate that it is better suited for handling DTDGs.