Independence-promoted disentangled dynamic graph attention network for out-of-distribution generalization

Dynamic graph neural networks (DyGNNs) have attracted increasing attention in recent years due to their strong ability to model graph structural and temporal dynamics. However, the existing DyGNNs fail to generalize under distribution shifts, which naturally exist in dynamic graphs, since the patterns exploited by DyGNNs may be variant with respect to labels under distribution shifts. These distribution shifts lead to substantial performance degradation of existing DyGNNs when applied to out-of-distribution (OOD) settings. Specifically, the existing DyGNNs tend to overfit to spurious correlations between variant patterns and labels, making it difficult to capture invariant patterns that are stable under distribution shifts. As a result, current DyGNNs struggle with OOD generalization. To address these challenges, in this paper, we propose an independence-promoted disentangled dynamic graph attention network which can effectively capture invariant patterns in dynamic graphs and significantly improve OOD generalization performance under distribution shifts. First, we introduce a disentangled dynamic graph attention network that explicitly separates invariant patterns from variant ones. Second, we introduce a causality-inspired spatio-temporal intervention mechanism that generates diverse intervened distributions and minimizes the variance of predictions among these distributions to eliminate spurious correlations. Third, we design an independence-promoted disentanglement optimization framework based on invariance and independence regularizers to enhance the model's ability to capture invariant patterns for OOD generalization. We conduct extensive experiments on several real-world and synthetic datasets. The experimental results show that our method consistently outperforms state-of-the-art DyGNNs and OOD generalization baselines under various distribution shift scenarios. In addition, ablation studies and visualization analyses validate the effectiveness of each key module. We also highlight promising directions for future research on OOD generalization in DyGNNs.