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

Representation learning on continuous-time dynamic graphs (CTDGs) is critical for modeling evolving network behaviors. However, existing methods often fail to capture both temporal dynamics and structural nuances effectively. Since the community is well-known for manifesting the mesoscopic structure of graphs, we propose community-enhanced temporal walks (CTWalks), a novel framework that explicitly leverages community structures to enhance representation learning on CTDGs. CTWalks integrate three key innovations: 1) a community-guided temporal walk sampling strategy that captures intra and intercommunity interactions to mitigate locality bias, with theoretical guarantees; 2) a community-aware anonymization process that embeds contextual community labels for robust node representations; and 3) a neural ordinary differential equations-based encoding mechanism that models continuous temporal dynamics, including community information with high fidelity. Furthermore, we establish a theoretical connection between CTWalks and matrix factorization, revealing the principled foundation. Extensive experiments on six benchmark datasets, including the large-scale tgbl-comment dataset with approximately one million nodes, demonstrate that CTWalks significantly outperform ten state-of-the-art methods in temporal link prediction, achieving substantial improvements in Area Under the receiver operating characteristic Curve (AUC) and average precision (AP) scores across diverse settings. This work advances dynamic graph learning by bridging community-aware structural insights with continuous-time modeling, enabling more accurate and adaptable representations for real-world networks. Our implementation is publicly available at https://github.com/leonyuhe/CTWalks.