Research map / Label-efficient and robust learning / Noisy labels
ROG_PL: Robust Open-Set Graph Learning via Region-Based Prototype Learning
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 ROG_PL, a graph open-set node classification framework that combines similarity-based label propagation for denoising with region-based (K-means clustering) interior and border prototype learning, trained with classification and prototype diversity losses.
Problem
Open-set node classification on graphs where training data contains both in-distribution label noise (IND noise) and out-of-distribution noise (OOD noise), causing intra-class variety and inter-class confusion problems.
Approach
Iteratively denoise labels via kNN-graph label propagation and remove low-confidence nodes, then learn multiple open-set prototypes per class from non-overlapping K-means regions (interior prototypes from homogeneous clusters, border prototypes from mixed clusters), with an orthogonality diversity loss, and reject low-similarity test nodes as unknown.
Key results
- On Cora, Citeseer and Coauthor-CS with near OOD noise and 5%/25%/50% IND noise, ROG_PL reports the best F1 in most settings, e.g. 78.36 F1 / 91.00 AUROC on Cora at 5% IND noise.
- ROG_PL achieves an average 6.23% F1 and 6.62% AUROC improvement over the second-best method PNP on the near-OOD setting, per the authors.
- Under 5% far OOD noise (Pubmed-sourced), ROG_PL reports 77.62 F1 and 87.89 AUROC on Cora, outperforming PNP (72.21 / 79.51).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ROG_PL rather than the strongest listed baseline | Open-set node classification · Cora, 5% IND noise and near OOD noise | F1: 73.03 -> 78.36 | Helped |
| Use ROG_PL rather than the strongest listed baseline | Open-set node classification · Citeseer, 5% IND noise and near OOD noise | F1: 59.99 -> 62.24 | Helped |
| Use ROG_PL rather than the strongest listed baseline | Open-set node classification · Coauthor-CS, 5% IND noise and near OOD noise | F1: 83.56 -> 81.68 | Hurt |
| Add the k-nearest neighbor graph for label propagation (ablation) | ROG_PL · Cora, 5% IND noise and near OOD noise | F1: 74.87 -> 78.36 | Helped |
| Add label-propagation-based denoising (ablation) | ROG_PL · Cora, 5% IND noise and near OOD noise | AUROC: 86.56 -> 91.00 | Helped |
| Add prototype diversity loss L_div (ablation) | ROG_PL · Cora, 5% IND noise and near OOD noise | F1: 73.67 -> 78.36 | Helped |
| Use region-based clustering and border prototypes (ablation) | ROG_PL · Cora, 5% IND noise and near OOD noise | AUROC: 84.11 -> 91.00 | Helped |
Surprising findings
- Several closed-set GCN variants with softmax output yield AUROC 'fail' since they lack an unknown score, and the plain GCN_soft_tau baseline is competitive on some settings.
Limitations
- Performance drops sharply once IND noise exceeds a certain threshold (beyond about 50%).
- Evaluated only on three citation graphs with synthetic noise injection and a GCN backbone.
- Hyperparameters (threshold, k) selected by grid search on the test-related setup; code only promised 'available online'.
Experimental setup
Datasets: Cora Citeseer Coauthor-CS Pubmed
Benchmarks: Cora Citeseer Coauthor-CS
Metrics: macro-F1 AUROC accuracy
Baselines: GCN_soft GCN_sig GCN_soft_tau GCN_sig_tau NWR_tau Openmax OpenWGL G2Pxy NGC PNP
Builds on: GCN label propagation NWR OpenWGL PNP
Closest papers by meaning
- Structural Entropy Guided Unsupervised Graph Out-Of-Distribution Detection (AAAI 2025 · similarity 0.87)
- Subgraph Aggregation for Out-of-Distribution Generalization on Graphs (AAAI 2025 · similarity 0.86)
- Unknown-Aware Graph Regularization for Robust Semi-supervised Learning from Uncurated Data (AAAI 2024 · similarity 0.86)
- Harnessing Feature Resonance under Arbitrary Target Alignment for Out-of-Distribution Node Detection (NeurIPS 2025 · similarity 0.86)
- ML-GOOD: Towards Multi-Label Graph Out-Of-Distribution Detection (AAAI 2025 · similarity 0.86)
- Dissecting the Failure of Invariant Learning on Graphs (NeurIPS 2024 · similarity 0.86)
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.