Research map / Learning theory and algorithms / Clustering
Incomplete Multi-view Clustering via Hierarchical Semantic Alignment and Cooperative Completion
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 HSACC, a deep incomplete multi-view clustering framework that aligns views via mutual information and MMD-based adaptive weighting in a dual-level latent space, and jointly learns cross-view completion with clustering.
Problem
Incomplete multi-view clustering where some views are entirely missing for some samples; existing deep methods use static fusion or two-stage completion-then-clustering pipelines that propagate errors.
Approach
Two semantic levels: low-level mutual-information alignment across views, and high-level adaptive view weights from MMD-based distributional affinity between each view and an initial fused representation, followed by weighted fusion; missing views are implicitly recovered by cross-view MLP inference and optimized jointly with reconstruction and clustering.
Key results
- On Caltech101-20 at missing rate 0.5, HSACC improves ACC and ARI by 5.3% and 8.57% over the second-best method.
- On Noisy MNIST, HSACC's ACC drops only 6.92% from missing rate 0.3 to 0.7, while ICMVC drops 35.19%.
- Ablation on Caltech101-20 shows ACC rising from 42.07% (reconstruction only) to 76.81% with all four losses.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HSACC instead of the strongest reported baseline | Incomplete multi-view clustering · LandUse_21 missing rate 0.5 | ACC: 27.34 -> 28.11 | Helped |
| Use HSACC instead of the strongest reported baseline | Incomplete multi-view clustering · Noisy MNIST missing rate 0.5 | ACC: 92.34 -> 92.55 | Helped |
| Use HSACC instead of the strongest reported baseline | Incomplete multi-view clustering · Caltech101-20 missing rate 0.5 | ACC: 71.51 -> 76.81 | Helped |
| Use HSACC instead of the strongest reported baseline | Incomplete multi-view clustering · Hdigit missing rate 0.5 | ACC: 94.40 -> 95.57 | Helped |
| Use HSACC instead of the strongest reported baseline | Incomplete multi-view clustering · 100leaves missing rate 0.5 | ACC: 51.39 -> 52.17 | Helped |
| Add cross-view consistency loss L_MMI (ablation) | HSACC · Caltech101-20 | ACC: 42.07 -> 68.69 | Helped |
| Add distribution alignment loss L_MMD (ablation) | HSACC · Caltech101-20 | ACC: 68.69 -> 71.29 | Helped |
| Add inference consistency loss L_INF (ablation) | HSACC · Caltech101-20 | ACC: 71.29 -> 76.81 | Helped |
| Use all four loss components rather than partial loss combinations (ablation) | HSACC · Caltech101-20 | ACC: 72.31 -> 76.81 | Helped |
| Vary the loss weights lambda_1, lambda_2, lambda_3, and lambda_4 (ablation) | HSACC · Caltech101-20 missing rate 0.5 | clustering performance | No clear effect |
| Increase the number of views from 2 to 5 (ablation) | HSACC on Mfeat · Mfeat | ACC: 74.96 -> 88.64 | Helped |
Surprising findings
- Reported results are without error bars or statistical significance, since baselines did not report them.
Limitations
- No error bars or significance tests reported.
- Evaluated only on two-view settings for main results; multi-view scalability shown on a single Mfeat experiment.
- Limited discussion of computational cost and of the mechanism's sensitivity beyond a few hyperparameter sweeps.
Experimental setup
Datasets: LandUse_21 Noisy MNIST Caltech101-20 Hdigit 100leaves Mfeat
Benchmarks: LandUse_21 Noisy MNIST Caltech101-20 Hdigit 100leaves Mfeat
Metrics: ACC NMI ARI
Baselines: RPCIC MRL_CAL ICMVC MCAC PROLMP DCP DSIMVC SURE COMP
Builds on: contrastive learning maximum mean discrepancy domain adaptation
Closest papers by meaning
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.92)
- Collaborative Similarity Fusion and Consistency Recovery for Incomplete Multi-view Clustering (AAAI 2025 · similarity 0.91)
- Adaptive Feature Imputation with Latent Graph for Deep Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.90)
- Towards Resource-friendly, Extensible and Stable Incomplete Multi-view Clustering (ICML 2024 · similarity 0.90)
- Deep Variational Incomplete Multi-View Clustering: Exploring Shared Clustering Structures (AAAI 2024 · similarity 0.90)
- Sample-Level Cross-View Similarity Learning for Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.89)
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.