Research map / Graph learning / Knowledge and temporal graphs
Future Link Prediction Without Memory or Aggregation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CRAFT, a memory-free and aggregation-free temporal graph model for future link prediction that combines learnable node embeddings with cross-attention between each candidate destination and the source's recent neighbors.
Problem
Future link prediction on temporal graphs must generalize to both seen (repeated) and unseen (novel) edges, but existing memory- and aggregation-based models perform poorly on unseen edges.
Approach
Discard memory and aggregation modules and instead use learnable node embeddings as unique identifiers plus cross-attention where each candidate destination attends to the source's recent interactions (target-aware matching), trained with BPR loss and elapsed/repeat time encodings.
Key results
- CRAFT achieves best MRR on 7 of 8 unseen-dominant datasets, e.g., 91.72% on tgbl-comment vs 76.17% for GraphMixer and 75.48% on WikiLink vs 62.94% for TGN.
- CRAFT-R achieves best MRR on seen-dominant datasets such as mooc (62.32% vs 42.20% DyGFormer) and lastfm (55.21% vs 46.50%), with small deficits (<2%) on three datasets.
- On Yelp and WikiLink, CRAFT's inference time is under half that of the most efficient baseline JODIE; on WikiLink CRAFT trains an epoch in 0.6 hours.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CRAFT instead of the strongest listed baseline | Future link prediction · unseen-dominant datasets: ML-20M, Taobao, Yelp, Flickr, WikiLink, tgbl-review, tgbl-comment | MRR (%): 49.69%; 17.25%; 4.02%; 25.74%; 19.92%; 0.82%; 20.41% relative improvements, respectively | Helped |
| Use CRAFT instead of the strongest listed baseline | Future link prediction · GoogleLocal | MRR (%): -0.85% | Hurt |
| Use CRAFT instead of the strongest listed baseline | Future link prediction · YouTube | MRR (%): 0.08% | Helped |
| Use CRAFT-R instead of the strongest listed baseline | Future link prediction · reddit, mooc, lastfm, tgbl-coin, tgbl-flights | MRR (%): 0.71%; 47.68%; 18.73%; 17.07%; 29.63% relative improvements, respectively | Helped |
| Use CRAFT-R instead of the strongest listed baseline | Future link prediction · wikipedia, uci, Flights | MRR (%): -0.50%; -1.96%; -1.15% | Hurt |
| Add learnable node embeddings | TGAT and DyGFormer · GoogleLocal, Yelp, tgbl-review | MRR | Helped |
| Use destination-to-source-neighbor cross-attention instead of independent or neighborhood matching | Future link prediction models · GoogleLocal, Yelp, tgbl-review | MRR | Helped |
| Remove positional encoding (ablation) | CRAFT · MOOC and GoogleLocal | MRR | Hurt |
| Remove elapsed time encoding (ablation) | CRAFT · wikipedia, mooc, uci, tgbl-review, GoogleLocal | MRR | Hurt |
| Remove repeat time encoding (ablation) | CRAFT-R · wikipedia and uci | MRR | Hurt |
| Replace positional encoding with time encoding (ablation) | CRAFT · wikipedia and mooc | MRR | Hurt |
| Replace BPR loss with BCE loss (ablation) | CRAFT · Five evaluated datasets | MRR | No clear effect |
Surprising findings
- Simply adding learnable node embeddings to TGAT and DyGFormer yields large gains, e.g., up to fivefold on Yelp for TGAT-LE.
Limitations
- Designed for transductive future link prediction; cold-start/inductive cases with little history are not handled.
- Does not use node or edge features in the current version.
- Not applicable to other tasks such as dynamic node classification without extension.
Experimental setup
Datasets: ML-20M Taobao Yelp GoogleLocal Flickr YouTube WikiLink tgbl-review tgbl-coin tgbl-comment tgbl-flights wikipedia
Benchmarks: TGB-Seq TGB wikipedia reddit mooc lastfm uci Flights
Metrics: MRR
Baselines: JODIE DyRep TGAT TGN CAWN GraphMixer DyGFormer SGNN-HN SASRec
Builds on: DyGFormer DyGLib TGB-Seq TGN
Closest papers by meaning
- TAMI: Taming Heterogeneity in Temporal Interactions for Temporal Graph Link Prediction (NeurIPS 2025 · similarity 0.90)
- Improving Temporal Link Prediction via Temporal Walk Matrix Projection (NeurIPS 2024 · similarity 0.89)
- Using Time-Aware Graph Neural Networks to Predict Temporal Centralities in Dynamic Graphs (NeurIPS 2024 · similarity 0.87)
- Revisiting Link Prediction: a data perspective (ICLR 2024 · similarity 0.87)
- Analytical Construction on Geometric Architectures: Transitioning from Static to Temporal Link Prediction (ICML 2025 · similarity 0.87)
- GNNs Meet Sequence Models Along the Shortest-Path: an Expressive Method for Link Prediction (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.