Research map / Graph learning / Graph self-supervised learning
AtlasULP: Domain-aware Universal Link Prediction via Relation Atlas
Yujing Liu, Yixin Liu, Yu Zheng, Lianhua Chi, Alan Wee-Chung Liew, Hengtao Shen, Shirui Pan
Read the paperCheck your idea against 137,217 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 AtlasULP, a domain-aware universal link prediction approach built on a relation atlas, aiming for training-free link prediction on arbitrary unseen graphs.
Problem
Conventional link prediction requires dataset-specific training for each graph, incurring high computational and maintenance costs; existing universal link prediction explores training-free inference on unseen graphs.
Approach
Introduces a domain-aware universal link prediction framework using a relation atlas to enable training-free inference across arbitrary unseen graphs (details beyond the abstract not available).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace UniLP with AtlasULP (relation atlas + domain prompting) | UniLP · 7 unseen graphs | Hits@50: 68.00 -> 71.69 | Improved |
| remove entire Local Refinement from Relation Atlas (ablation) | AtlasULP · Celegans, USAir, NS, Facebook | Hits@50: average 10.5% drop when removed | Improved |
| add Global Perception (Chebyshev multi-order path responses) to Relation Atlas (ablation) | AtlasULP · Celegans, USAir, NS, Facebook | Hits@50 | Improved |
| add local connectivity patterns (L1) or structural role similarity (L2) (ablation) | AtlasULP · Celegans, USAir, NS, Facebook | Hits@50 | Improved |
| add domain-aware prompt learning (ablation) | AtlasULP · ULP model ablation | Hits@50 | Improved |
| add deterministic structural sampling of context (ablation) | AtlasULP · ULP model ablation | Hits@50 | Improved |
| add in-context scoring with support sets (ablation) | AtlasULP · ULP model ablation | Hits@50 | Improved |
| compute pair features by sparse statistics instead of ego-subgraph extraction (UniLP) | AtlasULP · Facebook | running time: 1862.31s -> 29.47s | Improved |
| increase global perception hops K_gp (ablation) | AtlasULP · parameter sensitivity | Hits@50 | Improved |
| set local refinement hops K_lr to 3 (vs smaller or larger) (ablation) | AtlasULP · parameter sensitivity | Hits@50 | Mixed |
| increase support set size beyond 20 (ablation) | AtlasULP · parameter sensitivity | Hits@50 | No clear effect |
| scale training data and model size | AtlasULP · scaling analysis | Hits@50: exceeds 70% in large-data, large-model regime | Improved |
Closest papers by meaning
- Revisiting Link Prediction: a data perspective (ICLR 2024 · similarity 0.88)
- Future Link Prediction Without Memory or Aggregation (NeurIPS 2025 · similarity 0.87)
- GNNs Meet Sequence Models Along the Shortest-Path: an Expressive Method for Link Prediction (NeurIPS 2026 · similarity 0.86)
- A Topological Perspective on Demystifying GNN-Based Link Prediction Performance (ICLR 2024 · similarity 0.86)
- Pure Message Passing Can Estimate Common Neighbor for Link Prediction (NeurIPS 2024 · similarity 0.85)
- Bridging Theory and Practice in Link Representation with Graph Neural Networks (NeurIPS 2025 · similarity 0.85)
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.
