Community-Enhanced Temporal Walks: Debiasing Locality Representation Learning on Continuous-Time Dynamic Graphs.

Area of Science:

• Graph Neural Networks
• Network Science
• Machine Learning

Background:

• Representation learning on continuous-time dynamic graphs (CTDGs) is crucial for understanding evolving networks.
• Existing methods struggle to effectively capture both temporal dynamics and complex graph structures.
• Graphs often exhibit community structures that are vital for revealing mesoscopic properties.

Purpose of the Study:

• To propose a novel framework, CTWalks, that leverages community structures for enhanced representation learning on CTDGs.
• To improve the modeling of temporal dynamics and structural nuances in evolving networks.
• To advance the accuracy and adaptability of representations for real-world dynamic networks.

Main Methods:

• Community-guided temporal walk sampling to capture intra- and inter-community interactions, mitigating locality bias.
• Community-aware anonymization embedding contextual community labels for robust node representations.
• Neural ordinary differential equations (NODE) for high-fidelity modeling of continuous temporal dynamics and community information.

Main Results:

• CTWalks significantly outperform ten state-of-the-art methods in temporal link prediction across six benchmark datasets.
• Substantial improvements in Area Under the receiver operating characteristic Curve (AUC) and average precision (AP) scores were achieved.
• Demonstrated superior performance on the large-scale tgbl-comment dataset with approximately one million nodes.

Conclusions:

• CTWalks effectively bridge community-aware structural insights with continuous-time modeling for dynamic graph learning.
• The framework provides more accurate and adaptable representations for complex, evolving real-world networks.
• The theoretical connection to matrix factorization provides a principled foundation for the proposed methods.