Research map / Learning theory and algorithms / Clustering

Robust Multi-view Clustering against Imperfect Information

NeurIPS 2026 · Learning theory and algorithms › Clustering · approach: Incomplete multi-view clustering · 0 citations (OpenAlex)

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 PLCI, a unified multi-view clustering framework that treats the cross-view counterpart as a latent variable and infers it from instance reliability plus prototype-level optimal transport, handling both incomplete views and noisy correspondences.

Problem

Multi-view clustering under imperfect information, where views are missing (incomplete views) and cross-view correspondences are mismatched (noisy correspondences) simultaneously; existing IV- or NC-oriented methods handle only one.

Approach

Model each anchor's target-view counterpart as a latent variable whose posterior combines observed counterparts weighted by GMM-estimated correspondence reliability (memorization effect) with a prototype-level semantic counterpart obtained by reliability-weighted optimal transport across views.

Key results

• Across five feature-based datasets at IIR=0.8, PLCI averages 22.07 ACC / 32.28 NMI / 38.83 ACC-avg ... (reported averages: ACC 38.83, NMI 32.28, ARI 22.07) and best average at all IIR settings versus 10 baselines.
• On SUN RGB-D with raw RGB and depth inputs (ResNet-18), PLCI improves over the second-best method at IIR=0.8 by 4.19 ACC, 3.99 NMI, and 2.07 ARI.
• Two-stage combinations of DIVIDE and CorreGen perform clearly worse than PLCI, e.g., at IIR=0.8 on Scene15 PLCI reaches 37.78 ACC vs. 30.47 for the best ordering.

Surprising findings

• Sequentially cascading IV-oriented and NC-oriented methods (D+C, C+D) performs poorly and order-dependent, underperforming the unified approach.
• SURE and CAMERA evaluated on clean data (IIR=0 style) are reported as strong baselines, showing the baselines' dependence on correspondence quality.

Limitations

• Built on top of DIVIDE, so gains depend on that backbone; generalizability shown only on Completer with a single dataset.
• Hyperparameters such as alpha=0.2 and k_top=3 are fixed without a detailed sensitivity analysis in the main text.
• Imperfect information is simulated via random masking and shuffling rather than real-world noise.

Experimental setup

Datasets: Scene15 LandUse21 Reuters CCV20 HandWritten SUN RGB-D

Benchmarks: Scene15 LandUse21 Reuters CCV20 HandWritten SUN RGB-D

Metrics: ACC NMI ARI MSE

Baselines: SURE CAMERA GCFAgg DIVIDE FreeCSL CANDY ROLL CorreGen Completer

Builds on: DIVIDE CorreGen memorization effect of deep neural networks contrastive loss

Closest papers by meaning

• Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.89)
• Incomplete Multi-view Clustering via Hierarchical Semantic Alignment and Cooperative Completion (NeurIPS 2025 · similarity 0.89)
• AdaptCMVC: Robust Adaption to Incremental Views in Continual Multi-view Clustering (CVPR 2025 · similarity 0.89)
• Investigating and Mitigating the Side Effects of Noisy Views for Self-Supervised Clustering Algorithms in Practical Multi-View Scenarios (CVPR 2024 · similarity 0.88)
• Medusa: A Multi-Scale High-order Contrastive Dual-Diffusion Approach for Multi-View Clustering (CVPR 2025 · similarity 0.88)
• Deep Variational Incomplete Multi-View Clustering: Exploring Shared Clustering Structures (AAAI 2024 · similarity 0.88)