Handling distribution shifts on dynamic graphs via causal invariance principles.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Dynamic Graph Neural Networks (DyGNNs) often fail due to violated independent and identically distributed assumptions in real-world evolving graph structures.
• Distribution shifts in dynamic graphs present a significant challenge for the generalization of existing deep learning models.

Purpose of the Study:

• To propose DCIP (Distribution Shifts on Dynamic Graphs via Causal Invariance Principles) to address generalization issues caused by distribution shifts in dynamic graphs.
• To uncover and leverage causal patterns that remain stable across different environments within dynamic graph data.

Main Methods:

• Developed a multi-feature extraction module using interaction frequency coding to capture implicit node interaction patterns.
• Designed a frequency-domain causal disentanglement architecture combining Fourier Transform and Transformer to separate causal from non-causal graph components.
• Introduced a virtual intervention regularization strategy to enforce the stability of learned causal modes by perturbing non-causal elements.

Main Results:

• DCIP demonstrated consistent outperformance over existing methods across multiple tasks on six dynamic graph datasets.
• Experiments included validation on four distinct distribution shift datasets, confirming DCIP's robustness.
• The proposed method effectively handles distribution shifts by learning stable causal patterns.

Conclusions:

• DCIP offers a robust solution for improving the generalization of DyGNNs in the presence of distribution shifts.
• The causal invariance principles effectively stabilize learned representations across evolving graph environments.
• The method provides a promising direction for developing more reliable graph neural networks in dynamic, real-world scenarios.