Research map / Graph learning / Graph algorithms and clustering
One Node One Model: Featuring the Missing-Half for Graph Clustering
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 FPGC, a feature-personalized graph clustering method that builds a per-node model via a squeeze-and-excitation feature selector and adds a feature-cross augmentation for contrastive learning.
Problem
Attributed graph clustering methods rely mainly on topology and neglect node features, especially cluster-relevant personalized features and low-order feature interactions; high-dimensional features make feature selection for clustering hard.
Approach
"One node one model" paradigm: each node's label is a combination of predictions from cluster-group models, with weights from a squeeze-and-excitation block that selects top-n cluster-relevant features; a feature-cross augmentation (pairwise products of field-wise filtered features) provides low-order interactions for contrastive learning. The method is an SGC-based model made efficient via a Hada
Key results
- On Cora, FPGC reaches 79.19% ACC and 59.55% NMI, beating the runner-up by 4.04% ACC and 2.65% NMI.
- On EAT, FPGC reports 58.90% ACC vs 58.35% for CONVERT and 58.35 region, with 15.54% ACC gain over DyFSS per the text.
- FPGC achieves best reported ACC/NMI on all 7 tabled datasets in the main clustering table, and is faster than CGC and DyFSS on large graphs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add feature-cross augmentation instead of omitting augmentation (ablation) | FPGC · Cora | ACC: 75.95 -> 79.19 | Helped |
| replace classical graph augmentation with feature-cross augmentation (ablation) | FPGC · Cora | ACC: 76.42 -> 79.19 | Helped |
| add node-personalized feature gating g(X^n) (ablation) | FPGC · Cora | ACC: 78.52 -> 79.19 | Helped |
| add node-personalized feature gating g(X^n) (ablation) | FPGC · Pubmed | ACC: 71.20 -> 71.03 | Hurt |
| replace SGC with DAGNN as the base model (ablation) | FPGC · Cora | ACC: 79.19 -> 78.27 | Hurt |
| replace SGC with DAGNN as the base model (ablation) | FPGC · AMAP | ACC: 78.44 -> 79.52 | Helped |
| use features with the top 33.3% highest excitation scores (ablation) | FPGC feature selection · Cora and EAT | ACC | Helped |
| use feature-cross augmentation (ablation) | FPGC · Citeseer | ACC: 71.04 -> 72.59 | Helped |
| use feature-cross augmentation (ablation) | FPGC · Pubmed | ACC: 66.98 -> 71.03 | Helped |
| use feature-cross augmentation (ablation) | FPGC · UAT | ACC: 57.25 -> 58.07 | Helped |
| use feature-cross augmentation (ablation) | FPGC · AMAP | ACC: 77.45 -> 78.44 | Helped |
| use feature-cross augmentation (ablation) | FPGC · EAT | ACC: 57.82 -> 58.90 | Helped |
Surprising findings
- FPGC does not beat the baselines on Pubmed, which the authors attribute to large node count but only 3 clusters.
- Without the feature-selection gate g(X^n), performance degrades in most but not all cases, and feature cross augmentation alone gives only modest gains over alternative graph augmentations on some datasets.
Limitations
- Performance on Pubmed lags, attributed to the small number of clusters relative to node count.
- Hyperparameters (k, m, n, lambda) are set per dataset via grid search, with the dataset-specific settings tuned to the reported results.
Experimental setup
Datasets: Cora CiteSeer Pubmed AMAP UAT EAT BAT Flickr Twitch-Gamers
Benchmarks: Cora CiteSeer Pubmed AMAP UAT EAT BAT Flickr Twitch-Gamers
Metrics: ACC NMI running time
Baselines: SSGC MVGRL SDCN DFCN AGE MCGC FGC CGC CONVERT SCGC CCGC DGCN
Builds on: SGC DAGNN APPNP squeeze-and-excitation networks DGCN
Closest papers by meaning
- THESAURUS: Contrastive Graph Clustering by Swapping Fused Gromov-Wasserstein Couplings (AAAI 2025 · similarity 0.85)
- On the Expressivity and Sample Complexity of Node-Individualized Graph Neural Networks (NeurIPS 2024 · similarity 0.84)
- Node Identifiers: Compact, Discrete Representations for Efficient Graph Learning (ICLR 2025 · similarity 0.83)
- FUG: Feature-Universal Graph Contrastive Pre-training for Graphs with Diverse Node Features (NeurIPS 2024 · similarity 0.83)
- Local Graph Clustering with Noisy Labels (ICLR 2024 · similarity 0.83)
- Union Subgraph Neural Networks (AAAI 2024 · similarity 0.83)
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.