Research map / Graph learning / GNN architectures
Evaluating Graph Neural Networks for Link Prediction: Current Pitfalls and New Benchmarking
Juanhui Li, Harry Shomer, Haitao Mao, Shenglai Zeng, Yao Ma, Neil Shah, Jiliang Tang, Dawei Yin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Replace global random evaluation negatives with per-positive corruptions selected using RA, PPR, and feature similarity | Link prediction evaluation · Cora, Citeseer, Pubmed, ogbl-collab, ogbl-ddi, ogbl-ppa, ogbl-citation2 | mean model standard deviation: 5.19 -> 0.79; 5.94 -> 0.88; 4.14 -> 0.35; 1.49 -> 0.96; 2.13 -> 0.36; 6.77 -> 3.49; 1.39 -> 0.59 | Improved |
| Use heuristic-selected, personalized hard negatives instead of existing randomly sampled negatives | Link prediction evaluation · Most datasets; especially Cora, Citeseer, Pubmed, and ogbl-citation2 | MRR | Hurt |
| Use HeaRT negatives rather than existing negatives | Link prediction evaluation · ogbl-ppa | MRR | Mixed |
| Convert ogbl-citation2 to an undirected graph when computing heuristic scores | Heuristic methods · ogbl-citation2 | MRR: around 50% -> around 75% | Improved |
| Include validation edges in the graph during testing | Neo-GNN · ogbl-collab | Hits@50: 57.52 -> 66.13 | Improved |
| Train BUDDY with one negative per positive instead of six | BUDDY · ogbl-ddi | Hits@20: 78.51 -> 29.60 | Hurt |
| Remove pretraining of the GNN used to initialize Neo-GNN | Neo-GNN · ogbl-ddi | Hits@20: 63.57 -> 20.95 | Hurt |
| Use ranking-based metrics rather than relying on AUC alone | Link prediction evaluation · ogbl-ppa | Hits@K: 0; >99% | Mixed |
Closest papers by meaning
- A Topological Perspective on Demystifying GNN-Based Link Prediction Performance (ICLR 2024 · similarity 0.91)
- Implicit degree bias in the link prediction task (ICML 2025 · similarity 0.91)
- Revisiting Link Prediction: a data perspective (ICLR 2024 · similarity 0.89)
- On the Impact of Feature Heterophily on Link Prediction with Graph Neural Networks (NeurIPS 2024 · similarity 0.89)
- Bridging Theory and Practice in Link Representation with Graph Neural Networks (NeurIPS 2025 · similarity 0.89)
- NetInfoF Framework: Measuring and Exploiting Network Usable Information (ICLR 2024 · similarity 0.89)
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.
