Research map / Graph learning / Graph self-supervised learning
Towards Understanding Parametric Generalized Category Discovery on Graphs
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.
Introduces Graph Generalized Category Discovery (GGCD), gives a Wasserstein-based theoretical bound for parametric GCD, shows popular graph contrastive learning violates the required conditions, and proposes SWIRL, a prototype-based dilation contrastive loss that outperforms adapted VGCD baselines on synthetic and real graphs.
Problem
Understanding when and how known (old) classes help generalized category discovery on graphs with parametric classifiers, where existing GCD methods were designed for non-graph image data and lack theoretical foundations.
Approach
Quantify old-new class relationships via a joint (embedding, label) Wasserstein distance to derive a provable GCD loss upper bound; use a Pairwise Markov Random Field view to show InfoNCE/SupCon graph contrastive learning yields random category embedding relations; propose SWIRL, which uses prototype-based differentiated repulsion forces between old and new clusters to satisfy the derived lessons.
Key results
- SWIRL achieves the highest HRScore on Cora (64.53 vs 62.40 for Vanilla GCD), Citeseer-free overall leader in most datasets, and the best HRScore on A-Computers (61.46 vs 56.88 for Vanilla GCD).
- On synthetic CSBM graph D3, SWIRL reaches HRScore 94.03 versus 64.76 for SimGCD.
- On Citeseer, SWIRL (55.62 HRScore) trails Vanilla GCD (65.35) but beats SimGCD (46.36) and UNO+ (41.48).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SWIRL instead of the strongest HRScore baseline | GGCD methods · Cora | HRScore: 62.40±6.59 -> 64.53±4.63 | Helped |
| Use SWIRL instead of the strongest HRScore baseline | GGCD methods · Citeseer | HRScore: 65.35±6.10 -> 55.62±3.03 | Hurt |
| Use SWIRL instead of the strongest HRScore baseline | GGCD methods · Wiki | HRScore: 51.32±1.27 -> 52.93±2.72 | Helped |
| Use SWIRL instead of the strongest HRScore baseline | GGCD methods · A-Photo | HRScore: 76.72±5.12 -> 77.62±4.27 | Helped |
| Use SWIRL instead of the strongest HRScore baseline | GGCD methods · A-Computers | HRScore: 56.88±3.79 -> 61.46±4.54 | Helped |
| Replace SimGCD with SWIRL using a GCN encoder | D1 · D1 | HRScore: 68.22 -> 96.24 | Helped |
| Replace SimGCD with SWIRL using a GCN encoder | D3 · D3 | HRScore: 64.76 -> 94.03 | Helped |
| Replace SWIRL's GCN encoder with GPR (ablation) | SWIRL · D1 | HRScore: 96.24 -> 87.72 | Hurt |
| Replace SWIRL's GCN encoder with GPR (ablation) | SWIRL · D3 | HRScore: 94.03 -> 82.39 | Hurt |
| Add entropy regularization (ablation) | MLP on Dz2 · Dz2 | New RACC: 0 -> 72 | Helped |
| Add entropy regularization (ablation) | MLP on Dz2 · Dz2 | Reject ACC: 50 -> 98 | Helped |
| Replace SimGCD's GCN encoder with GPR | SimGCD · D3 | HRScore: 64.76 -> 88.77 | Helped |
Surprising findings
- SWIRL with a GPR encoder underperforms SimGCD-GPR on the D3 synthetic graph, which the authors attribute to weak intra-cluster cohesion.
- Standard InfoNCE/SupCon graph contrastive learning yields random category embedding distances across seeds, so the PMRF analysis suggests such losses cannot control inter-class relations.
Limitations
- SWIRL with GPR encoders underperforms on some datasets, since GPR weakens intra-cluster cohesion.
- Theoretical analysis relies on assumptions (bounded embeddings, Lipschitz classifier, Lipschitz transferable classifier) that are only partly justified.
- Real-world results are mixed; SWIRL trails Vanilla GCD on Citeseer and some other datasets.
Experimental setup
Datasets: Cora Citeseer Wiki A-Photo A-Computers
Benchmarks: Cora Citeseer Wiki A-Photo A-Computers
Metrics: HRScore Old RACC New RACC Reject ACC
Baselines: GCN SS-KM UNO+ Vanilla GCD SimGCD SimGCD-GPR
Builds on: GRACE SimGCD Vanilla GCD (SS-KM) UNO+ InfoNCE
Closest papers by meaning
- Federated Generalized Category Discovery (CVPR 2024 · similarity 0.89)
- Prior-Constrained Association Learning for Fine-Grained Generalized Category Discovery (AAAI 2025 · similarity 0.89)
- Generalized Category Discovery via Reciprocal Learning and Class-Wise Distribution Regularization (ICML 2025 · similarity 0.89)
- Active Generalized Category Discovery (CVPR 2024 · similarity 0.89)
- Generalized Category Discovery under Domain Shift: A Frequency Domain Perspective (NeurIPS 2025 · similarity 0.88)
- DebGCD: Debiased Learning with Distribution Guidance for Generalized Category Discovery (ICLR 2025 · similarity 0.88)
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.