Zening Li, Rong-Hua Li, Yichun Yang, Meihao Liao · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840579
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Node representation learning on dynamic graphs is a fundamental task for numerous real-world applications. However, existing dynamic Graph Neural Networks (GNNs) face two critical challenges: (1) the inability to effectively model crucial directional information inherent in many real-world interactions, and (2) suffer from poor scalability due to the neighbor explosion problem. To address these limitations, we propose ID2GL, a novel and scalable framework for incremental node representation learning on dynamic directed graphs. ID2GL features a decoupled architecture and introduces a novel propagation mechanism based on complex-valued personalized PageRank. This mechanism not only captures directionality via complex phases but also incorporates multi-hop neighbor information, which effectively alleviates the neighbor explosion issue. To handle continuous updates efficiently, we propose an incremental update mechanism that requires only local computations on affected nodes, supported by theoretical guarantees on the post-update representation error. Finally, a transformer-based prediction module is utilized to capture the temporal evolution of node representations for various downstream tasks. Extensive experiments on multiple real-world datasets, including the massive Papers100M dataset, demonstrate that ID2GL significantly outperforms state-of-the-art baselines in both node classification and link prediction tasks, and achieves competitive efficiency and scalability.
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
No comments yet — start the discussion below.