Research map / Graph learning / Graph self-supervised learning
Topology-Aware Dynamic Reweighting for Distribution Shifts on Graph
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 TAR, a minimax node-reweighting method that runs Wasserstein gradient flow on the graph (geometric Wasserstein space) to adjust node weights for distribution shifts, and evaluates on GOOD node-OOD datasets and long-tailed citation benchmarks.
Problem
Node classification with GNNs degrades when training and test nodes come from different distributions (covariate or concept shift), and existing invariant learning and Group DRO reweighting methods either rely on unverified invariance assumptions or ignore graph topology.
Approach
Solve the inner maximization over node densities by time-discretized gradient flow in a discrete geometric Wasserstein space, so weight changes propagate only along graph edges, with entropy and topology penalties; the one-step flow is shown equal to a worst-case distribution in a local uncertainty set.
Key results
- On GOOD benchmarks TAR achieves best results in 6 of 8 settings, e.g. CBAS covariate 79.86% vs 78.71% best baseline and Twitch concept ROC-AUC 49.32%.
- On PubMed-LT, TAR reaches 67.33% bAcc vs 57.83% for Re-Weight and 57.77% for ERM (with ERM backbone), about 10 points improvement.
- Ablation with reconnected graph (TAR-N) performs worse than TAR on most settings, supporting preserving the original graph structure.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TAR instead of ERM | GCN · WebKB, CBAS, Twitch, and Cora OOD shifts | accuracy or ROC-AUC | Helped |
| Use TAR instead of ERM | GAT on Cora-LT, CiteSeer-LT, and PubMed-LT · class-imbalanced node classification | Acc, bAcc, and F1 | Helped |
| Use TAR instead of GraphENS | GraphENS backbone · CiteSeer-LT and PubMed-LT | Acc, bAcc, and F1 | Helped |
| Use TAR instead of GraphENS | GraphENS backbone · Cora-LT | Acc and bAcc | Hurt |
| Reconnect labeled nodes with added edges (TAR-N) instead of preserving the original graph (ablation) | TAR · WebKB, CBAS, Twitch, Cora, and long-tailed datasets | accuracy or ROC-AUC | Mixed |
| Increase gradient-flow iterations; performance saturates at 10 on WebKB and CBAS, and 30 on Twitch (ablation) | TAR · WebKB, CBAS, and Twitch concept shift | accuracy or ROC-AUC | Helped |
| Set sample-weight smoothness parameter β to 0.01 (ablation) | TAR · OOD concept-shift datasets | accuracy or ROC-AUC | Helped |
Surprising findings
- Most graph OOD baselines fail to beat ERM across all datasets, and DANN is only on par with ERM under concept shift on Twitch and WebKB.
- EERM performs best on WebKB covariate shift but fails under concept shift and runs out of memory on Twitch and Cora.
Limitations
- Unlabeled nodes between sampled training nodes break gradient flow connectivity; the mean-loss imputation heuristic is admitted to be suboptimal.
- Theoretical error bound constant C is only stated to depend on loss, beta and N, and the bound relies on prior results.
- Hyperparameters (T_in, beta, learning rate) are tuned per setting via grid search, and the OOD and imbalance experiments use different backbones.
Experimental setup
Datasets: WebKB CBAS Twitch Cora-Word (GOOD) Cora-LT CiteSeer-LT PubMed-LT
Benchmarks: GOOD WebKB GOOD CBAS GOOD Twitch GOOD Cora Cora-LT CiteSeer-LT PubMed-LT
Metrics: accuracy ROC-AUC balanced accuracy F1
Baselines: ERM IRM VREx Group DRO DANN Deep Coral EERM SR-GCN Re-Weight CB Loss TAM GraphENS
Builds on: Group DRO discrete geometric Wasserstein distance Wasserstein gradient flow (Liu et al.) GOOD benchmark GraphENS
Closest papers by meaning
- From Distribution to Geometry: Stable Graph Generalization via Invariant Barycenters (ICML 2026 · similarity 0.90)
- Learning to Reweight for Generalizable Graph Neural Network (AAAI 2024 · similarity 0.90)
- Handling Distribution Shifts on Graphs: An Invariance Perspective (ICLR 2022 · similarity 0.90)
- Dissecting the Failure of Invariant Learning on Graphs (NeurIPS 2024 · similarity 0.88)
- Class-Imbalanced Graph Learning without Class Rebalancing (ICML 2024 · similarity 0.87)
- Dynamic Rescaling for Training GNNs (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.