DyFiLM: A framework to handle the distribution shifts on dynamic graphs with hypernetworks.

Dynamic graph representation learning has garnered increasing attention since dynamic graphs can accurately reflect the changes and evolutionary processes in the real world. Existing approaches typically train a fixed model to capture temporal patterns and then utilize the fixed model to infer future evolution. They assume that the dynamic graph evolves with the same law over time. While the underlying data generation distribution of graphs may shift over time and introduce new evolution patterns. To resolve this challenge, we propose a learning-to-learn framework entitled Dynamic Feature-wise Linear Modulation (DyFiLM), which employs a hypermodel to adjust the representation learning model to change with the distribution shifts. By employing a hypermodel to directly modulate the representation learning model based on time-varying input data, our framework captures evolutionary patterns from diverse time and expresses them through the modulation. Training both hypermodel and representation learning model in a distribution-shifting environment endows the framework with the capability for cross-distribution generalization. We apply our proposed framework to three different models and conduct extensive experiments on four datasets to verify the effectiveness of DyFiLM. The experimental results demonstrate that the DyFiLM achieves significant improvements compared with related approaches.