Erfan Moshiri, Amin Beheshti, M. A. Hakim Newton, Min Li, Pooyan Asgari · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3839975
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).
Dynamic graph learning faces key trade-offs: self-attention suffers from quadratic computational scaling, while memory-based methods compress interaction histories into fixed vectors, losing fine-grained temporal patterns. Furthermore, existing link prediction models rely on destination nodes as attention anchors, making them inadequate for standalone node classification. We introduce DANCE, a memory-free framework using learnable embeddings or global tokens as self-contained queries to perform identity-aware trajectory matching via linear-time asymmetric cross-attention over raw neighbor histories. This eliminates the quadratic bottleneck while preserving full per-node expressive capacity. Enriched with multi-scale temporal features, DANCE achieves state-of-the-art AUC-ROC performance (up to 6.2% relative improvement over best baselines) with significantly lower training time and memory overhead. To handle cold-start settings, our two inductive variants substitute node-specific parameters with global tokens, ensuring robust generalization on unseen nodes. Code is available on GitHub (https://github.com/erfanmoshiri/DANCE).
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
No comments yet — start the discussion below.