Research map / Graph learning / Graph self-supervised learning
Evaluating Robustness and Uncertainty of Graph Models Under Structural Distributional Shifts
Gleb Bazhenov, Denis Kuznedelev, Andrey Malinin, Artem Babenko, Liudmila Prokhorenkova
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 |
|---|---|---|---|
| Test ERM on nodes selected by locality rather than popularity | ERM · Structural shifts, average across datasets | Accuracy drop (ID to OOD): -13.56\,\% locality vs -4.39\,\% popularity | Hurt |
| Test SE on density-based rather than popularity-based shifts | SE · Structural shifts, average across datasets | AUROC: 81.40 vs 52.10 | Hurt |
| Use Mixup instead of ERM | Graph node classification · Popularity and locality structural shifts | Accuracy rank on OOD nodes: 2.1 (Popularity); 2.4 (Locality) | Improved |
| Use DE instead of other evaluated methods | Graph node classification · Density structural shift, OOD nodes | Accuracy rank: 2.0 | Improved |
| Use DE uncertainty estimates instead of other evaluated methods | OOD detection · Locality and density structural shifts | AUROC rank: 1.1 (Locality); 2.7 (Density) | Improved |
| Add the modified architecture (MLP preprocessing, skip connections, fewer layers, and SAGE) to ERM (ablation) | ERM · Popularity and density shifts, OOD nodes | Performance versus baseline win/tie/loss: 5/2/0 | Improved |
| Add the modified architecture (MLP preprocessing, skip connections, fewer layers, and SAGE) to ERM (ablation) | ERM · Locality shift, OOD nodes | Performance versus baseline win/tie/loss: 3/2/2 | Mixed |
| Add the modified architecture (MLP preprocessing, skip connections, fewer layers, and SAGE) to SE (ablation) | SE · Popularity and locality shifts | Performance versus baseline win/tie/loss: 0/0/7 (Popularity); 0/1/6 (Locality) | Hurt |
| Add the modified architecture (MLP preprocessing, skip connections, fewer layers, and SAGE) to SE (ablation) | SE · Density shift | Performance versus baseline win/tie/loss: 4/0/3 | Mixed |
| Increase the ID share from 70% to 90% (ablation) | ERM · Locality shift | Average accuracy drop (ID to OOD): -4.65\,\% -> -1.02\,\% | Improved |
| Increase the ID share from 70% to 90% (ablation) | SE · Locality shift | Average AUROC: 78.71 -> 72.95 | Hurt |
Closest papers by meaning
- Quantifying Distributional Invariance in Causal Subgraph for IRM-Free Graph Generalization (NeurIPS 2025 · similarity 0.88)
- Matcha: Mitigating Graph Structure Shifts with Test-Time Adaptation (ICLR 2025 · similarity 0.87)
- Topology-Aware Dynamic Reweighting for Distribution Shifts on Graph (ICML 2025 · similarity 0.87)
- Distribution Learning Meets Graph Structure Sampling (NeurIPS 2025 · similarity 0.87)
- No Metric to Rule Them All: Toward Principled Evaluations of Graph-Learning Datasets (ICML 2025 · similarity 0.87)
- GraphMETRO: Mitigating Complex Graph Distribution Shifts via Mixture of Aligned Experts (NeurIPS 2024 · similarity 0.87)
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.
