Research map / Graph learning / Graph self-supervised learning
Concorde: Geometry-Aware Link Prediction via Decoupled Energy Minimization
Ege Demirci, Ambuj K Singh
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 Concorde, an energy-based link prediction framework that decomposes link likelihood into three decoupled branches, including a parameter-free spectral anchor, to address spectral instability in sparse graphs.
Problem
Link prediction in sparse graphs, where diffusion heuristics lack connectivity and geometric methods suffer spectral instability when coupling learned representations with diffusion operators.
Approach
Decomposes link likelihood into three structurally decoupled branches within an energy-based framework, with a parameter-free spectral anchor as one branch, avoiding instability from coupling learned representations with diffusion operators.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add mixed-curvature residual branch (ablation) | Concorde · CiteSeer | AP: 2.59 -> 5.05 | Improved |
| add mixed-curvature residual branch (ablation) | Concorde · Power | AP: 2.72 -> 6.39 | Improved |
| add non-local geodesic steering field (ablation) | Concorde · Power | AP: 4.17 -> 6.39 | Improved |
| add non-local geodesic steering field (ablation) | Concorde · OGBL-DDI | AP: 38.38 -> 42.32 | Improved |
| add residual and field branches on top of invariant anchor (ablation) | Concorde · OGBL-DDI | AP: 28.70 -> 42.32 | Improved |
| use product manifold H x S instead of Euclidean latent (ablation) | Concorde residual · OGBL-DDI | AP: 38.52 -> 42.32 | Improved |
| use product manifold H x S instead of Euclidean latent (ablation) | Concorde residual · Power | AP: 5.85 -> 6.39 | Improved |
| include hyperbolic component (vs spherical only) (ablation) | Concorde residual · Power | AP: 5.78 -> 6.39 | Improved |
| include spherical component (vs hyperbolic only) (ablation) | Concorde residual · OGBL-DDI | AP: 38.32 -> 42.32 | Improved |
| add feature-aware Ollivier-Ricci curvature proxy (ablation) | Concorde residual · OGBL-DDI | AP: 39.22 -> 42.32 | Improved |
| add local neighborhood geometry statistics (ablation) | Concorde residual · Power | AP: 4.85 -> 6.39 | Improved |
| add geometry on top of same-loss learned reweighting (ablation) | Concorde · OGBL-DDI | AP: 38.15 -> 42.32 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Gelato — Gelato: Graph Edit Distance via Autoregressive Neural Combinatorial Optimization (ICLR 2026) · better on 6 of 7 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Analytical Construction on Geometric Architectures: Transitioning from Static to Temporal Link Prediction (ICML 2025 · similarity 0.86)
- Complete-Tree Space Favors Data-Efficient Link Prediction (ICML 2025 · similarity 0.86)
- GRAPHLCP: Structure-Aware Localized Conformal Prediction on Graphs (NeurIPS 2026 · similarity 0.86)
- Distilling Graph Geometry: Knowledge Gap from GNNs to MLPs (NeurIPS 2026 · similarity 0.86)
- A Topological Perspective on Demystifying GNN-Based Link Prediction Performance (ICLR 2024 · similarity 0.86)
- Two Heads Are Better Than One: Boosting Graph Sparse Training via Semantic and Topological Awareness (ICML 2024 · 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.
