Research map / Graph learning / GNN architectures
Mixture of Link Predictors on Graphs
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 Link-MoE, a mixture-of-experts model that routes each node pair to existing link predictors (GNN4LP models and heuristics-based experts) using a gating network over pairwise heuristic and feature inputs, trained in a two-step procedure.
Problem
Link prediction methods typically apply one pairwise strategy uniformly to all node pairs, while different pairs in the same dataset need different heuristics and different GNN4LP models perform best in different regimes.
Approach
Treat existing link predictors as experts and learn a gating network, conditioned on node-pair features and structural heuristics (CN, AA, RA, shortest path, Katz, PPR, degree), that assigns per-pair weights to experts; experts are trained first, then only the gating model is trained on a split of the validation set.
Key results
- Link-MoE achieves 18.71% relative MRR improvement over the best baseline on Pubmed (53.10 vs. about 44.73 for NBFNet) and 9.59% Hits@100 improvement on ogbl-ppa (69.39).
- Link-MoE reaches 44.03 MRR on Cora and 67.49 MRR on Citeseer, the best reported results in the main table on those datasets (Cora gain 10.80%).
- On ogbl-collab Link-MoE scores 71.32 Hits@50, improving 4.67% over the second-best method, and on ogbl-citation2 it gives 91.25 MRR, a 0.58% gain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use pair-specific gating to combine experts instead of LPFormer | Link-MoE · Cora | MRR: 10.80% | Helped |
| use pair-specific gating to combine experts instead of LPFormer | Link-MoE · Citeseer | MRR: 3.16% | Helped |
| use pair-specific gating to combine experts instead of NBFNet | Link-MoE · Pubmed | MRR: 18.71% | Helped |
| use pair-specific gating to combine experts instead of LPFormer | Link-MoE · ogbl-collab | Hits@50: 4.67% | Helped |
| use pair-specific gating to combine experts instead of LPFormer | Link-MoE · ogbl-ppa | Hits@100: 9.59% | Helped |
| use pair-specific gating to combine experts instead of Global-Ensemble | Link-MoE · ogbl-citation2 | MRR: 0.58% | Helped |
| activate only the Top-3 experts instead of all experts (ablation) | Link-MoE · ogbl-collab | Hits@50: 66.13 -> 71.94 | Helped |
| use 3 experts instead of all experts (ablation) | Link-MoE · ogbl-collab | Hits@50: 71.32 -> 71.25 | No clear effect |
| use 4 experts instead of all experts (ablation) | Link-MoE · ogbl-collab | Hits@50: 71.32 -> 72.75 | Helped |
| activate only the Top-3 experts instead of all experts (ablation) | Link-MoE · Pubmed | MRR: 53.10 -> 51.13 | Hurt |
| use 3 experts instead of all experts (ablation) | Link-MoE · Pubmed | MRR: 53.10 -> 52.03 | Hurt |
| use 4 experts instead of all experts (ablation) | Link-MoE · Pubmed | MRR: 53.10 -> 52.30 | Hurt |
Surprising findings
- Simply ensembling heuristics via an MLP fails to outperform GNN4LP methods, indicating that naive combination of heuristics is insufficient.
- Mean-Ensemble and Global-Ensemble often underperform single best experts, showing that uniform or global expert weighting is inadequate.
Limitations
- Gating model is trained on a portion of the original validation set, which gives Link-MoE extra labeled data relative to baselines; the fairness argument is informal.
- NBFNet and PEG are excluded from OGB experts due to out-of-memory issues, and some experts (e.g., SEAL) are computationally expensive.
- Authors note Link-MoE does not always achieve the best results on Cora and Citeseer, attributed to the small validation sets.
Experimental setup
Datasets: Cora Citeseer Pubmed ogbl-collab ogbl-ppa ogbl-citation2 Chameleon Squirrel
Benchmarks: Cora Citeseer Pubmed ogbl-collab ogbl-ppa ogbl-citation2 Chameleon Squirrel
Metrics: MRR Hits@K
Baselines: CN AA RA Shortest Path Katz Node2Vec MF MLP GCN GAT SAGE GAE
Builds on: Mixture of Experts NCN NCNC Neo-GNN BUDDY LPFormer
Closest papers by meaning
- GNNs Meet Sequence Models Along the Shortest-Path: an Expressive Method for Link Prediction (NeurIPS 2026 · similarity 0.88)
- A Topological Perspective on Demystifying GNN-Based Link Prediction Performance (ICLR 2024 · similarity 0.87)
- Revisiting Link Prediction: a data perspective (ICLR 2024 · similarity 0.87)
- Bridging Theory and Practice in Link Representation with Graph Neural Networks (NeurIPS 2025 · similarity 0.86)
- On the Impact of Feature Heterophily on Link Prediction with Graph Neural Networks (NeurIPS 2024 · similarity 0.86)
- Implicit degree bias in the link prediction task (ICML 2025 · 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.