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

Area of Science:

• Machine Learning
• Graph Neural Networks
• Dynamic Systems

Background:

• Dynamic graphs model real-world evolving systems.
• Current methods assume static evolution laws, failing with distribution shifts.
• Adapting to changing graph patterns is crucial for accurate predictions.

Purpose of the Study:

• Introduce a learning-to-learn framework, Dynamic Feature-wise Linear Modulation (DyFiLM).
• Enable representation learning models to adapt to evolving graph data distributions.
• Enhance cross-distribution generalization capabilities.

Main Methods:

• DyFiLM employs a hypermodel to modulate a core representation learning model.
• The hypermodel adapts the learning model based on time-varying input data.
• Joint training in distribution-shifting environments fosters adaptability.

Main Results:

• DyFiLM significantly improves performance across three different base models.
• Extensive experiments on four datasets validate the framework's effectiveness.
• The approach successfully captures and expresses diverse temporal evolutionary patterns.

Conclusions:

• DyFiLM provides a robust solution for dynamic graph representation learning under distribution shifts.
• The framework demonstrates superior performance compared to existing methods.
• DyFiLM enhances the adaptability and generalization of dynamic graph models.