Handling distribution shifts on dynamic graphs via causal invariance principles.

Dynamic Graph Neural Networks (DyGNNs) typically assume that training and test data follow an independent and identically distributed assumption. However, in real-world scenarios, the evolving nature of dynamic graph structures violates the assumption, causing classic deep learning models to fail to generalize to test data. In this paper, we propose DCIP to handle Distribution Shifts on Dynamic Graphs via Causal Invariance Principles, which aims to uncover causal patterns that remain stable across environments and thereby addressing distribution shifts. Specifically, we first develop a multi-feature extraction module to explore implicit node interaction patterns via interaction frequency coding. Additionally, we design a frequency-domain causal disentanglement architecture that combines Fourier Transform and Transformer to separate causal patterns from non-causal patterns. Finally, we introduce a virtual intervention regularization strategy that actively perturbs non-causal components to generate a set of intervention distributions, thereby enforcing the stability of the learned causal modes across shifting environments. Extensive experiments on six dynamic graph datasets and four distribution shift datasets demonstrate that DCIP consistently outperforms existing methods in multiple tasks. The code is publicly available at https://github.com/zhabng/DCIP.