Research map / Graph learning / GNN architectures
Structural Information Enhanced Graph Representation for Link Prediction
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 SIEG, a link prediction framework that removes neighbor node features from a GNN (using DRNL structural labels) and adds a Binary Structural Transformer that attends only between target nodes with pairwise heuristic encodings.
Problem
GNN-based link prediction is node-centric: it misses long-range pairwise structure between target nodes and neighbor node features may add noise for link prediction.
Approach
Remove neighbor node features so the GNN encodes only structural (DRNL) labels within enclosing subgraphs, and add Binary Structural Attention that restricts attention to the two target nodes and injects pairwise heuristics (SPD, Jaccard, Adamic-Adar) into attention logits.
Key results
- SIEG w/o NF reaches 63.22% Hits@100 on ogbl-ppa, ranking first over SEAL-style and NGNN baselines (NGNN 59.71%).
- SIEG w/o NF achieves 89.87% MRR on ogbl-citation2 (NGNN 88.91%) and 98.74% AUC on Pubmed, ranking first on these.
- Removing BST drops ogbl-ppa Hits@100 to 60.05% and Pubmed AUC to 97.94%, showing BST contributes pairwise structure features missed by GNN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SIEG w/o NF instead of Neural CN | Link prediction · ogbl-ppa | Hits@100 (%): 61.19±0.85 -> 63.22±1.74 | Helped |
| Use SIEG w/o NF instead of NGNN | Link prediction · ogbl-citation2 | MRR (%): 88.91±0.22 -> 89.87±0.18 | Helped |
| Use SIEG w/o NF instead of SUREL | Link prediction · ogbl-vessel | AUC (%): 84.96±0.68 -> 96.87±0.01 | Helped |
| Use SIEG w/o NF instead of PS2 | Link prediction · Pubmed | AUC (%): 97.60±0.10 -> 98.74±0.04 | Helped |
| Use SIEG w/o NF instead of S-VGAE | Link prediction · Cora | AUC (%): 94.1±0.1 -> 93.92±0.05 | Hurt |
| Use SIEG w/o NF instead of S-VGAE | Link prediction · Citeseer | AUC (%): 94.7±0.2 -> 92.10±0.44 | Hurt |
| Add neighbor node features (ablation) | SIEG · ogbl-ppa | Hits@100 (%): 63.22±1.74 -> 61.88±4.33 | Hurt |
| Add neighbor node features (ablation) | SIEG · ogbl-citation2 | MRR (%): 89.87±0.18 -> 89.57±0.10 | Hurt |
| Add neighbor node features (ablation) | SIEG · ogbl-vessel | AUC (%): 96.87±0.01 -> 96.77±0.01 | Hurt |
| Add neighbor node features (ablation) | SIEG · Pubmed | AUC (%): 98.74±0.04 -> 98.50±0.09 | Hurt |
| Add neighbor node features (ablation) | SIEG · Cora | AUC (%): 93.92±0.05 -> 93.04±0.54 | Hurt |
| Add neighbor node features (ablation) | SIEG · Citeseer | AUC (%): 92.10±0.44 -> 90.85±0.74 | Hurt |
Surprising findings
- Removing neighbor node features improved SEAL on ogbl-ppa, ogbl-citation2 and ogbl-vessel, suggesting neighbor features can be noisy for link prediction.
- Replacing BST with a fully-connected Transformer lowered performance across variants, which the authors attribute to information from nodes away from targets.
Limitations
- Fully-connected Transformer variants cannot run on OGB datasets due to quadratic complexity, so the BST-vs-FCT comparison is limited to small graphs.
- Gains on some classic datasets are small or absent (e.g., Cora and Citeseer with neighbor features are below baselines), and the ogbl-ppa variance is large (±4.33 for w/ NF).
Experimental setup
Datasets: ogbl-ppa ogbl-citation2 ogbl-vessel Pubmed Cora Citeseer
Benchmarks: ogbl-ppa ogbl-citation2 ogbl-vessel Pubmed Cora Citeseer
Metrics: Hits@100 MRR AUC
Baselines: CN AA RA Node2vec MF MLP GCN GraphSAGE SEAL NGNN SUREL Neural CN
Builds on: SEAL NGNN SortPooling Graphormer-style topological encoding GAE
Closest papers by meaning
- Pure Message Passing Can Estimate Common Neighbor for Link Prediction (NeurIPS 2024 · similarity 0.90)
- GNNs Meet Sequence Models Along the Shortest-Path: an Expressive Method for Link Prediction (NeurIPS 2026 · similarity 0.90)
- A Topological Perspective on Demystifying GNN-Based Link Prediction Performance (ICLR 2024 · similarity 0.89)
- Open Your Eyes: Vision Enhances Message Passing Neural Networks in Link Prediction (ICML 2025 · similarity 0.89)
- Structural Entropy Based Graph Structure Learning for Node Classification (AAAI 2024 · similarity 0.89)
- NLGT: Neighborhood-based and Label-enhanced Graph Transformer Framework for Node Classification (AAAI 2025 · 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.