Research map / Graph learning / GNN architectures
Deep Contrastive Graph Learning with Clustering-Oriented Guidance
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 DCGL, a deep unsupervised graph clustering model that combines a pseudo-siamese GCN/auto-encoder with feature-level and cluster-level contrastive learning guided by centroids, learning local and global graphs for general (non-graph) data clustering.
Problem
Graph-based clustering of general data without a prior graph relies heavily on a low-quality initial graph, causing representation collapse and lacking clustering-oriented guidance during training.
Approach
Pairs a GCN branch with an auto-encoder branch in a pseudo-siamese network, uses k-means centroids to select negatives in feature-level contrastive learning, builds a local (adaptive KNN) and global (personalized PageRank diffusion) graph, and aligns them in cluster space via cluster-level contrastive learning.
Key results
- DCGL reports ACC 81.93 and NMI 86.18 on YaleB, versus 67.02 and 80.68 for the best competing method AdaGAE.
- On PIE DCGL reaches ACC 99.86 and NMI 99.87, above AdaGAE's 97.97 and 98.18.
- Ablation on TOX-171 shows removing cluster-level contrastive learning (DCGL-w/C) drops ACC from 51.46 to 42.11.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DCGL instead of the strongest reported baseline | DCGL · TOX-171 | NMI: 30.60 -> 34.75 | Helped |
| Use DCGL instead of the strongest reported baseline | DCGL · ORL | ACC: 69.50 -> 75.50 | Helped |
| Use DCGL instead of the strongest reported baseline | DCGL · TR41 | ACC: 72.78 -> 73.23 | Helped |
| Use DCGL instead of the strongest reported baseline | DCGL · YaleB | ACC: 67.02 -> 81.93 | Helped |
| Use DCGL instead of the strongest reported baseline | DCGL · Isolet | ACC: 69.49 -> 71.73 | Helped |
| Use DCGL instead of the strongest reported baseline | DCGL · PIE | ACC: 97.97 -> 99.86 | Helped |
| Use DCGL instead of the strongest reported baseline | DCGL · USPS | ACC: 76.44 -> 78.32 | Helped |
| Add feature-level contrastive learning (ablation) | DCGL · TOX-171 | ACC: 47.95 -> 51.46 | Helped |
| Add cluster-level contrastive learning and siamese graph convolution (ablation) | DCGL · TOX-171 | ACC: 42.11 -> 51.46 | Helped |
| Add feature-level clustering guidance (ablation) | DCGL · four datasets shown in the figure | ACC | Helped |
| Add cluster-level clustering guidance (ablation) | DCGL · four datasets shown in the figure | ACC | Helped |
| Add both feature-level and cluster-level clustering guidance (ablation) | DCGL · four datasets shown in the figure | ACC | Helped |
Surprising findings
- Graph-based methods designed for graph-type data, DAEGC and DCRN, perform poorly on general clustering tasks, with DCRN near 5% NMI on USPS.
Limitations
- Relies on k-means centroids for negative sample selection, so rough clustering quality may affect contrastive guidance.
- Evaluated only on seven small-to-medium benchmark datasets.
Experimental setup
Datasets: TOX-171 ORL TR41 YaleB Isolet PIE USPS
Benchmarks: TOX-171 ORL TR41 YaleB Isolet PIE USPS
Metrics: ACC NMI
Baselines: k-means STSC CAN PCAN RGC GEMMF DAEGC DCRN DEC VGAE SpectralNet SUBLIME
Builds on: SDCN AdaGAE SUBLIME InfoNCE Personalized PageRank graph diffusion
Closest papers by meaning
- Cluster-Guided Contrastive Graph Clustering Network (AAAI 2023 · similarity 0.93)
- GLCC: A General Framework for Graph-Level Clustering (AAAI 2023 · similarity 0.92)
- Global Graph Propagation with Hierarchical Information Transfer for Incomplete Contrastive Multi-view Clustering (AAAI 2025 · similarity 0.89)
- THESAURUS: Contrastive Graph Clustering by Swapping Fused Gromov-Wasserstein Couplings (AAAI 2025 · similarity 0.88)
- Multi-View Graph Clustering via Node-Guided Contrastive Encoding (ICML 2025 · similarity 0.88)
- S3GCL: Spectral, Swift, Spatial Graph Contrastive Learning (ICML 2024 · similarity 0.87)
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.