Self-Supervised Temporal Graph Learning With Temporal and Structural Intensity Alignment.

Area of Science:

• Graph Machine Learning
• Network Science
• Data Mining

Background:

• Temporal graphs capture dynamic node interactions over time.
• Existing methods often overlook high-order structural information, limiting representation quality.
• Effective temporal graph learning requires integrating both temporal dynamics and structural patterns.

Purpose of the Study:

• To propose S2T, a self-supervised method for temporal graph learning.
• To enhance node representations by incorporating both temporal and high-order structural information.
• To improve performance on graph-based tasks with dynamic data.

Main Methods:

• S2T combines first-order temporal information with high-order structural information.
• It calculates two conditional intensities using different combinations of temporal and structural data.
• An alignment loss optimizes node representations by minimizing the difference between these intensities.
• Structural information is considered at local (neighbor sequences) and global (all nodes) levels.

Main Results:

• The proposed S2T model achieves significant performance improvements.
• Experiments show up to a 10.13% increase in performance compared to state-of-the-art methods.
• S2T effectively extracts both temporal and structural features for richer node representations.

Conclusions:

• S2T offers a more informative approach to temporal graph learning.
• Integrating high-order structural information alongside temporal data is crucial for performance.
• The self-supervised S2T method demonstrates superior capabilities in handling dynamic graph data.