Dynamic Graph Representation Learning via Coupling-Process Model.

Area of Science:

• Machine Learning
• Graph Representation Learning
• Network Science

Background:

• Dynamic graph representation learning is crucial for analyzing evolving networks.
• Existing methods often overlook the distinct nature of different edge types.
• Real-world dynamic graphs exhibit diverse edge behaviors, categorized as infrequent evolutive and frequent interactive edges.

Purpose of the Study:

• To propose a novel model, DyCPM, that addresses the limitations of existing dynamic graph representation learning methods.
• To effectively capture the distinct dynamic mechanisms of evolutive and interactive edges.
• To improve the performance of dynamic link prediction by integrating topological and temporal information from different edge types.

Main Methods:

• Developed a coupling-process model (DyCPM) for dynamic graph representation learning.
• Designed a neural network parameterized discrete process for evolutive edges.
• Utilized a neural network parameterized temporal point process (TPP) for interactive edges.
• Implemented a coupling mechanism to integrate information from both processes via a shared embedding matrix.

Main Results:

• The proposed DyCPM model successfully aggregates structural and temporal information from both evolutive and interactive edges.
• The model generates low-dimensional node embeddings that capture the nuances of different edge types.
• Experimental evaluations on real-world datasets demonstrate superior performance compared to state-of-the-art baselines in dynamic link prediction.

Conclusions:

• DyCPM effectively models the distinct properties of evolutive and interactive edges in dynamic graphs.
• The coupling mechanism allows for synergistic integration of information from different edge types.
• The proposed approach significantly enhances dynamic link prediction accuracy by leveraging heterogeneous edge dynamics.