Research map / Graph learning / Graph algorithms and clustering
Attribute-Missing Graph Clustering Network
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 AMGC, a unified alternating clustering-and-imputation framework for deep graph clustering on attribute-missing graphs, using cluster-oriented imputation and a dual non-contrastive clustering loss.
Problem
Deep clustering on graphs where only a subset of nodes has complete attributes; existing two-stage impute-then-cluster methods decouple imputation from clustering and are not designed for clustering.
Approach
Alternates between a clustering step (K-means on EMA-encoded graph embeddings) and an imputation step where K-nearest neighbors within the same predicted cluster fill missing attributes, while a dual non-contrastive cluster-center loss (ProPos-inspired) separates clusters and an edge reconstruction loss trains the imputation encoder.
Key results
- On Co.CS, AMGC reaches 72.61% ACC, outperforming AGC-DRR and CCGC by 13.53% and 7.73% ACC.
- AMGC outperforms SAT, SVGA, and ITR by up to 31.04% ACC on Cora and 25.01% on Pubmed.
- On Co.Physics, most baselines run out of memory while AMGC achieves 77.68% ACC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the dual non-contrastive clustering loss (ablation) | AMGC · Cora | ACC: 57.17 -> 66.65 | Helped |
| add the cluster-oriented imputation scheme (ablation) | AMGC · Cora | ACC: 63.72 -> 66.65 | Helped |
| add the dual non-contrastive clustering loss (ablation) | AMGC · Citeseer | ACC: 53.90 -> 60.92 | Helped |
| add the cluster-oriented imputation scheme (ablation) | AMGC · Citeseer | ACC: 57.08 -> 60.92 | Helped |
| add the dual non-contrastive clustering loss (ablation) | AMGC · Pubmed | ACC: 43.02 -> 64.56 | Helped |
| add the cluster-oriented imputation scheme (ablation) | AMGC · Pubmed | ACC: 62.51 -> 64.56 | Helped |
| add the dual non-contrastive clustering loss (ablation) | AMGC · Co.CS | ACC: 56.44 -> 72.61 | Helped |
| add the cluster-oriented imputation scheme (ablation) | AMGC · Co.CS | ACC: 70.18 -> 72.61 | Helped |
| add the dual non-contrastive clustering loss (ablation) | AMGC · Co.Physics | ACC: 73.15 -> 77.68 | Helped |
| add the cluster-oriented imputation scheme (ablation) | AMGC · Co.Physics | ACC: 75.95 -> 77.68 | Helped |
Surprising findings
- SVGA shows more stable clustering results than AMGC on some datasets, which the authors attribute to its Gaussian latent assumption.
Limitations
- Experiments assume a fixed 40% attribute-complete split; the imputation relies on K-means cluster quality.
- Only five datasets evaluated; baselines are mostly attribute-complete methods adapted to the setting.
Experimental setup
Datasets: Cora Citeseer Pubmed Coauthor CS Coauthor Physics
Benchmarks: Cora Citeseer Pubmed Coauthor CS Coauthor Physics
Metrics: ACC NMI ARI F1
Baselines: SDCN GDCL AGCN AGC-DRR HSAN CCGC IMKAM DIMVC SAT SVGA ITR
Builds on: SDCN ProPos BGRL SAT ITR
Closest papers by meaning
- Adaptive Feature Propagation for Attribute-Missing Graph Clustering (NeurIPS 2026 · similarity 0.93)
- Attribute-Missing Multi-view Graph Clustering (CVPR 2025 · similarity 0.92)
- Adaptive Feature Imputation with Latent Graph for Deep Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.90)
- Scalable Attribute-Missing Graph Clustering via Neighborhood Differentiation (ICML 2025 · similarity 0.89)
- DGCLUSTER: A Neural Framework for Attributed Graph Clustering via Modularity Maximization (AAAI 2024 · similarity 0.86)
- Propagate and Inject: Revisiting Propagation-Based Feature Imputation for Graphs with Partially Observed Features (ICML 2025 · 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.