Research map / Graph learning / Graph self-supervised learning
Multi-View Graph Clustering via Node-Guided Contrastive Encoding
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 NGCE, a node-feature-guided contrastive graph encoder for multi-view graph clustering that unifies homophilic and heterophilic information, evaluated on six homophilic and heterophilic multi-view graph datasets.
Problem
Multi-view graph clustering with GNN-based contrastive learning struggles to integrate node features with graph structure and to handle both homophilic and heterophilic graphs.
Approach
Guides graph encoding with node features: a homophily-weighted sum of a node-feature similarity matrix and the adjacency matrix (weights from pseudo-labels), a masked-feature recovery contrastive loss through parameter-free GCNs, and a cross-view contrastive loss using cluster centroids and kNN neighbors as positives/negatives.
Key results
- On ACM, NGCE reaches 80.5% NMI, 85.0% ARI, 94.7% ACC and 94.8% F1, the best among compared methods.
- On Texas, NGCE reports 47.8% NMI and 77.6% ACC, substantially above the runner-up on heterophilous data.
- On Chameleon, NGCE reaches 22.6% NMI and 42.2% ACC; on IMDB and DBLP it is competitive but not uniformly best (e.g., DBLP NMI 79.1 vs BMGC 80.1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add reconstruction loss LRec (ablation) | NGCE · ACM | ACC%: 86.8 -> 94.7 | Helped |
| add reconstruction loss LRec (ablation) | NGCE · Chameleon | ACC%: 33.0 -> 42.2 | Helped |
| add contrastive learning loss LC (ablation) | NGCE · ACM | ACC%: 92.1 -> 94.7 | Helped |
| add contrastive learning loss LC (ablation) | NGCE · Chameleon | ACC%: 39.9 -> 42.2 | Helped |
| add contrastive loss LC1 (ablation) | NGCE · ACM | ACC%: 93.7 -> 94.7 | Helped |
| add contrastive loss LC1 (ablation) | NGCE · Chameleon | ACC%: 40.2 -> 42.2 | Helped |
| add contrastive loss LC2 (ablation) | NGCE · ACM | ACC%: 93.4 -> 94.7 | Helped |
| add contrastive loss LC2 (ablation) | NGCE · Chameleon | ACC%: 41.3 -> 42.2 | Helped |
| use the joint process instead of adjacency matrix A (ablation) | NGCE · ACM | ACC%: 80.5 -> 94.7 | Helped |
| use the joint process instead of adjacency matrix A (ablation) | NGCE · Chameleon | ACC%: 27.1 -> 42.2 | Helped |
| use refined node features Z instead of an affinity matrix built from original node features (ablation) | NGCE · ACM | ACC%: 89.8 -> 94.7 | Helped |
| use refined node features Z instead of an affinity matrix built from original node features (ablation) | NGCE · Chameleon | ACC%: 35.1 -> 42.2 | Helped |
Surprising findings
- On DBLP and IMDB, NGCE is not uniformly best; several baselines (e.g., MCGC, BMGC) match or exceed it on some metrics.
Limitations
- Depends on the availability of node features; not directly applicable to featureless graphs.
- Performance is sensitive to the GCN order hyperparameter, degrading sharply above order 15.
- Quadratic-in-N graph operations; scalability limited without anchor-graph approximations.
Experimental setup
Datasets: ACM DBLP IMDB Texas Chameleon Wisconsin
Benchmarks: ACM DBLP IMDB Texas Chameleon Wisconsin
Metrics: NMI ARI ACC F1
Baselines: VGAE DAEGC AGE O2MAC MvAGC AGCN MCGC DCRN DuaLGR CMGEC BMGC SMVC
Builds on: GraphMAE Denoising Autoencoder NT-Xent MVGRL DAEGC
Closest papers by meaning
- Global Graph Propagation with Hierarchical Information Transfer for Incomplete Contrastive Multi-view Clustering (AAAI 2025 · similarity 0.90)
- Medusa: A Multi-Scale High-order Contrastive Dual-Diffusion Approach for Multi-View Clustering (CVPR 2025 · similarity 0.89)
- EASEMVC:Efficient Dual Selection Mechanism for Deep Multi-View Clustering (CVPR 2025 · similarity 0.89)
- Homophily-Related: Adaptive Hybrid Graph Filter for Multi-View Graph Clustering (AAAI 2024 · similarity 0.88)
- THESAURUS: Contrastive Graph Clustering by Swapping Fused Gromov-Wasserstein Couplings (AAAI 2025 · similarity 0.88)
- A Hubness Perspective on Representation Learning for Graph-Based Multi-View Clustering (CVPR 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.