Research map / Graph learning / Knowledge and temporal graphs
Linkless Link Prediction via Relational Distillation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| relational KD (rank + distribution matching, LLP) vs direct logit/representation KD | MLP student · Transductive, 8 datasets | Hits@20: 10.59 points avg over direct KD | Helped |
| relational KD (LLP) vs teacher GNN | MLP student · Transductive, 8 datasets | Hits@20: outperforms GNN on 7 of 8 datasets; Photos -2.18 | Mixed |
| relational KD (LLP) vs MLP | MLP student · Production, non-OGB datasets | Hits@20: 12.01 points avg over MLP | Helped |
| relational KD (LLP) vs direct KD (logit and representation matching) | MLP student · Production, non-OGB datasets | Hits@20: 6.67 points avg | Helped |
| remove rank-based matching loss L_LLP_R (ablation) | LLP · Transductive Pubmed and CS | Hits@20: Pubmed 57.33 -> 55.35; CS 68.62 -> 66.61 | Helped |
| remove distribution-based matching loss L_LLP_D (ablation) | LLP · Transductive Pubmed and CS | Hits@20: Pubmed 57.33 -> 54.97; CS 68.62 -> 65.17 | Helped |
| remove rank-based matching loss L_LLP_R (ablation) | LLP · Production Pubmed and CS | Hits@20: Pubmed 53.48 -> 53.40; CS 60.74 -> 60.53 | No clear effect |
| remove distribution-based matching loss L_LLP_D (ablation) | LLP · Production Pubmed and CS | Hits@20: Pubmed 53.48 -> 48.58; CS 60.74 -> 60.13 | Helped |
| use distillation losses only (remove L_sup) with rank loss removed, i.e. distribution-only (ablation) | LLP · Transductive CS | Hits@20: CS 68.39 without L_sup and L_LLP_R vs 68.62 full; Pubmed 54.86 | Mixed |
| use distillation losses only (remove L_sup) in production setting (ablation) | LLP · Production Pubmed and CS | Hits@20: Pubmed 53.48 -> 39.35 (w/o rank, L_sup) and 41.43 (w/o distribution, L_sup); CS 60.74 -> 57.35 and 55.63 | Hurt |
| direct logit-matching KD (L_LM) and representation-matching KD (L_RM) vs plain MLP | MLP student · Transductive, 8 datasets | Hits@20: L_LM hurts on Cora (74.72 vs 78.06), Computers (12.81 vs 20.19); L_RM helps on CS (61.10 vs 34.01 MLP) | Mixed |
Closest papers by meaning
- GNNs Meet Sequence Models Along the Shortest-Path: an Expressive Method for Link Prediction (NeurIPS 2026 · similarity 0.90)
- Link Prediction Based on Graph Neural Networks (NeurIPS 2018 · similarity 0.90)
- A Topological Perspective on Demystifying GNN-Based Link Prediction Performance (ICLR 2024 · similarity 0.89)
- Bridging Theory and Practice in Link Representation with Graph Neural Networks (NeurIPS 2025 · similarity 0.88)
- Mixture of Link Predictors on Graphs (NeurIPS 2024 · similarity 0.88)
- Revisiting Link Prediction: a data perspective (ICLR 2024 · similarity 0.88)
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.