Title: TIDFormer: Leveraging Temporal and Interactive Dynamics for a Superior Dynamic Graph Transformer

Abstract:

Self-attention mechanisms (SAMs) have proven highly effective in capturing dependencies for sequence modeling, leading to the adoption of Transformer architectures in various dynamic graph neural networks (DGNNs). These models typically employ diverse encoding schemes to track the sequential evolution of dynamic graphs. However, the performance and computational efficiency of such Transformer-based DGNNs are inconsistent, underscoring the need to properly define SAMs for dynamic graphs and comprehensively encode temporal and interactive dynamics without relying on overly complex modules.

To address these challenges, we introduce TIDFormer, a dynamic graph Transformer designed to efficiently harness both Temporal and Interactive Dynamics. We clarify and validate the interpretability of our proposed SAM, resolving the issue of uninterpretable definitions found in prior research. To capture temporal dynamics, we utilize calendar-based time partitioning information. For interactive dynamics, we extract informative interaction embeddings for both bipartite and non-bipartite graphs by relying solely on sampled first-order neighbors. Additionally, we jointly model temporal and interactive features by capturing potential shifts in historical interaction patterns through a straightforward decomposition method.

We performed extensive experiments across multiple dynamic graph datasets to validate the effectiveness and efficiency of TIDFormer. The results show that TIDFormer outperforms state-of-the-art models in most datasets and experimental configurations. Moreover, TIDFormer demonstrates significant efficiency gains over previous Transformer-based approaches.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC