Answers are grounded in this paper and cite their sources.
Future link prediction on dynamic graphs must rank both recurring interactions and links between previously unconnected nodes. We extend the CRAFT-R candidate-aware cross-attention matcher with explicit historical recurrence statistics, time-weighted embedding similarity, and candidate-specific residual fusion. Under chronological splits and fixed 100-candidate queries on three competition datasets, we evaluate four retrained configurations alongside history-based ranking rules. The full model achieves test mean reciprocal ranks (MRRs) of 0.845, 0.296, and 0.576 on dataset1, dataset2, and dataset3, respectively, exceeding the backbone control by 0.033, 0.026, and 0.033. Removing the statistics branch reduces MRR by 0.020, 0.007, and 0.004; removing the similarity branch reduces it by 0.012, 0.021, and 0.012. Separate strict-past analyses show that direct recurrence is sensitive to historical coverage and to whether recent events remain visible at prediction time. The auxiliary signals therefore contribute differently across datasets, while the value of a longer or rolling history depends on the available evidence. The full model also leads seven evaluated baselines on AP, AUC, and MRR across the three datasets under the reported protocol.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: