Research map / Learning theory and algorithms / Clustering
Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding
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 ICMVC, an end-to-end incomplete multi-view clustering framework that transfers adjacency graphs across views, encodes incomplete data with GCNs, fuses views via instance-level attention, and uses contrastive and high-confidence self-training losses for joint clustering.
Problem
Incomplete multi-view clustering, where instances have missing views, and existing methods either ignore complementary information or treat missing-value handling and representation/clustering as separate stages.
Approach
Missing views' adjacency matrices are filled by transferring the graph from an observed view of the same instance; GCN message passing imputes missing representations; instance-level attention fusion plus a high-confidence target distribution (sharpened from max over view and fused assignments) guides an end-to-end clustering objective combined with instance-level and cluster-level contrastive los
Key results
- On MSRC-V1 at 30% missing rate, ICMVC reaches 87.72 ACC / 76.97 NMI / 74.03 ARI, beating the best baseline by about 10.96% ACC.
- On Noisy MNIST at 30% missing, ICMVC reaches 96.82 ACC versus 79.96 for Completer.
- Ablation on MSRC-V1 (eta=0.3): removing L_ins drops ACC from 87.72 to 80.38; removing L_hg drops it to 85.62.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ICMVC instead of the strongest reported baseline | Incomplete multi-view clustering · MSRC-V1 (30% missing) | ACC: 76.76 ± 6.37 -> 87.72 ± 0.47 | Helped |
| Use ICMVC instead of the strongest reported baseline | Incomplete multi-view clustering · Scene-15 (30% missing) | ACC: 31.90 ± 1.31 -> 36.20 ± 1.47 | Helped |
| Use ICMVC instead of the strongest reported baseline | Incomplete multi-view clustering · LandUse-21 (30% missing) | ACC: 24.80 ± 1.05 -> 26.94 ± 0.46 | Helped |
| Use ICMVC instead of the strongest reported baseline | Incomplete multi-view clustering · HandWritten (30% missing) | ACC: 76.61 ± 4.89 -> 82.34 ± 1.62 | Helped |
| Use ICMVC instead of the strongest reported baseline | Incomplete multi-view clustering · Noisy MNIST (30% missing) | ACC: 79.96 ± 2.23 -> 96.82 ± 0.11 | Helped |
| Add instance-level contrastive loss (ablation) | ICMVC · MSRC-V1 (30% missing) | ACC: 80.38 -> 87.72 | Helped |
| Add high-confidence guidance (ablation) | ICMVC · MSRC-V1 (30% missing) | ACC: 85.62 -> 87.72 | Helped |
| Add high-confidence guidance and clustering loss instead of using k-means (ablation) | ICMVC · MSRC-V1 (30% missing) | ACC: 70.67 -> 87.72 | Helped |
Surprising findings
- Performance degrades quickly when the missing rate exceeds 0.5, which the authors attribute to failure of consistency-based graph transfer.
- Several baselines (DSIMVC, DIMVC) perform poorly on some datasets, and deep baselines show large standard deviations.
Limitations
- Performance drops sharply at missing rates above 0.5.
- Computational complexity grows with the number of views.
- Loss terms are weighted equally without tuning trade-off parameters.
Experimental setup
Datasets: HandWritten Scene-15 LandUse-21 MSRC-V1 Noisy MNIST
Benchmarks: HandWritten Scene-15 LandUse-21 MSRC-V1 Noisy MNIST
Metrics: ACC NMI ARI
Baselines: BSV Concat PVC MIC DAIMC Completer DSIMVC DIMVC
Builds on: Contrastive Clustering DEC Completer DAIMC
Closest papers by meaning
- Incomplete Multi-view Clustering via Diffusion Contrastive Generation (AAAI 2025 · similarity 0.93)
- Adaptive Feature Imputation with Latent Graph for Deep Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.93)
- Incomplete Multi-view Clustering via Hierarchical Semantic Alignment and Cooperative Completion (NeurIPS 2025 · similarity 0.92)
- Deep Variational Incomplete Multi-View Clustering: Exploring Shared Clustering Structures (AAAI 2024 · similarity 0.92)
- Diffusion-based Missing-view Generation With the Application on Incomplete Multi-view Clustering (ICML 2024 · similarity 0.91)
- Incomplete and Unpaired Multi-View Graph Clustering with Cross-View Feature Fusion (AAAI 2025 · similarity 0.91)
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.