Research map / Learning theory and algorithms / Clustering
Imputation-free and Alignment-free: Incomplete Multi-view Clustering Driven by Consensus Semantic Learning
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 FreeCSL, an imputation- and alignment-free incomplete multi-view clustering framework that learns consensus prototypes in a shared semantic space via prototypical contrastive clustering, plus modularity-based graph clustering per view.
Problem
Incomplete multi-view clustering suffers from prototype shifts within views and semantic inconsistency across views due to missing data; existing methods rely on imputation and cross-view alignment that become unreliable at high missing rates.
Approach
Replace instance- and cluster-level consistency with a semantic-level paradigm: fuse view representations with completeness-based weights into consensus representations, derive consensus prototypes via k-means, and align all observations (paired or not) to them through swapped prototype contrast, while a modularity-based graph clustering loss enriches each view with cluster structure. No imputatio
Key results
- On ALOI-100 at missing rate 0.1, FreeCSL reaches 91.13% ACC, with reported ACC gains of 15.12%, 26.38%, 23.91%, 26.21% over the best baselines across r=0.1, 0.3, 0.5, 0.7.
- On Caltech-5V at r=0.7, FreeCSL achieves 83.64% ACC, and on NoisyMNIST at r=0.7 it reaches 92.19% ACC.
- Ablation shows the consensus semantic learning module contributes the most; on ALOI-100 ACC drops only 12.4% as r rises from 0.1 to 0.7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace CPM-Nets with FreeCSL | Caltech-5V · Caltech-5V, r = 0.7 | ACC (%): 79.50 -> 83.64 | Helped |
| Replace SURE with FreeCSL | ALOI-100 · ALOI-100, r = 0.7 | ACC (%): 50.23 -> 76.44 | Helped |
| Replace DIMVC with FreeCSL | YouTubeFace10 · YouTubeFace10, r = 0.7 | ACC (%): 74.94 -> 76.62 | Helped |
| Replace SURE with FreeCSL | NoisyMNIST · NoisyMNIST, r = 0.7 | ACC (%): 84.00 -> 92.19 | Helped |
| Add CSL and CSE to reconstruction-only FreeCSL (ablation) | FreeCSL · Caltech-5V, r = 0.7 | ACC (%): 51.93 -> 83.64 | Helped |
| Add CSL and CSE to reconstruction-only FreeCSL (ablation) | FreeCSL · ALOI-100, r = 0.7 | ACC (%): 23.78 -> 76.44 | Helped |
| Add CSE (ablation) | FreeCSL with REC + CSL · Caltech-5V, r = 0.7 | ACC (%): 79.14 -> 83.64 | Helped |
| Add CSE (ablation) | FreeCSL with REC + CSL · ALOI-100, r = 0.7 | ACC (%): 75.29 -> 76.44 | Helped |
| Add CSL (ablation) | FreeCSL with REC + CSE · Caltech-5V, r = 0.7 | ACC (%): 70.71 -> 83.64 | Helped |
| Add CSL (ablation) | FreeCSL with REC + CSE · ALOI-100, r = 0.7 | ACC (%): 32.73 -> 76.44 | Helped |
| Impute latent representations using cross-view K-NN graphs (ablation) | FreeCSL · Caltech-5V, r = 0.7 | ACC (%): 83.64 -> 77.86 | Hurt |
| Impute semantic representations using cross-view K-NN graphs (ablation) | FreeCSL · Caltech-5V, r = 0.7 | ACC (%): 83.64 -> 81.29 | Hurt |
Surprising findings
- Explicit imputation after consensus semantic learning (ISR) is competitive with FreeCSL on Caltech-5V, and ISR beats FreeCSL on some low-missing-rate ALOI-100 cases in the imputation comparison table.
Limitations
- Sensitive to the modularity regularizer weight lambda; large values let contrastive learning dominate graph clustering.
- Relies on k-means for prototypes and final prediction, and on hyperparameter choices tuned per dataset.
Experimental setup
Datasets: Caltech-5V ALOI-100 YoutubeFace10 NoisyMNIST
Benchmarks: Caltech-5V ALOI-100 YoutubeFace10 NoisyMNIST
Metrics: ACC NMI ARI
Baselines: CPM-Nets COMPLETER DIMVC SURE ProImp ICMVC DIVIDE
Builds on: contrastive clustering swapped prediction / knowledge distillation modularity maximization graph convolutional network
Closest papers by meaning
- Deep Variational Incomplete Multi-View Clustering: Exploring Shared Clustering Structures (AAAI 2024 · similarity 0.88)
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.88)
- Incomplete Multi-view Clustering via Hierarchical Semantic Alignment and Cooperative Completion (NeurIPS 2025 · similarity 0.88)
- Towards Resource-friendly, Extensible and Stable Incomplete Multi-view Clustering (ICML 2024 · similarity 0.87)
- Adaptive Feature Imputation with Latent Graph for Deep Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.87)
- Incomplete Multi-view Clustering via Diffusion Contrastive Generation (AAAI 2025 · 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.