EasyDGL: Encode, Train and Interpret for Continuous-Time Dynamic Graph Learning.

Area of Science:

• Graph Neural Networks
• Dynamic Systems Modeling
• Time Series Analysis

Background:

• Dynamic graphs are prevalent in real-world scenarios, necessitating flexible continuous-time modeling.
• Existing methods often lack interpretability or struggle with coupled spatiotemporal dynamics.

Purpose of the Study:

• To introduce EasyDGL, an easy-to-use pipeline for modeling dynamic graphs in continuous time.
• To enhance model fitting ability and interpretability for spatiotemporal graph dynamics.
• To provide a framework for quantifying the influence of learned frequency content on predictive tasks.

Main Methods:

• A temporal point process (TPP) modulated attention architecture for continuous-time graph dynamics with edge events.
• A principled loss function combining task-agnostic TPP posterior maximization and task-aware masking for dynamic graph prediction.
• Scalable perturbation-based analysis in the graph Fourier domain for model interpretability.

Main Results:

• EasyDGL demonstrates superior performance on time-conditioned predictive tasks across public benchmarks.
• The pipeline effectively quantifies the predictive power of frequency content learned from evolving graph data.
• Achieved strong fitting ability and interpretability in dynamic graph modeling.

Conclusions:

• EasyDGL offers a flexible and interpretable solution for continuous-time dynamic graph modeling.
• The framework advances the understanding of how models utilize frequency information in evolving graphs.
• Enables robust performance on tasks like dynamic link prediction and node classification.