Research map / Learning theory and algorithms / Clustering

Multi-view Granular-ball Contrastive Clustering

AAAI 2025 · Learning theory and algorithms › Clustering · approach: Incomplete multi-view clustering · 11 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 MGBCC, a deep multi-view clustering method that builds coarse granular balls (k-means-based) per view and applies contrastive learning between associated intra- and cross-view granular balls instead of individual instances or clusters.

Problem

Multi-view contrastive clustering suffers from false negatives at the instance level and overlooks local structure at the cluster level; the paper seeks multi-granularity contrastive learning that avoids both issues.

Approach

Segment each view's latent space into granular balls (k-means with k = N/p), link balls within a view by overlap and across views by intersection size of shared samples, and apply a contrastive loss over granular-ball centers so that negatives are non-associated balls rather than neighboring samples.

Key results

• On Cora, MGBCC reaches 65.44% ACC versus 49.07% for the best competing method (DealMVC).
• MGBCC achieves best or second-best ACC on most of the seven datasets, e.g., 95.77% ACC on BBCSport and 99.60% on MNIST-USPS.
• On Caltech101-20 ablation, granular-ball contrastive loss (p=2) raises ACC from 44.47% (instance-level, p=1) to 72.63%.

Measured effects reported in the paper

Surprising findings

• On ALOI-100, SURE (90.37% ACC) slightly outperforms MGBCC (88.39%), and on NoisyMNIST MGBCC ranks roughly third.
• Instance-level contrastive loss (p=1) gave only modest gains over reconstruction alone on Caltech101-20 (44.47% vs 40.28% ACC).

Limitations

• Performance degrades for large granularity parameter p, as larger balls cannot be associated reliably by overlap and intersection size.
• Not best on all datasets (e.g., ALOI-100 and NoisyMNIST), and some baseline numbers are quoted from original papers rather than rerun.
• Several hyperparameters (p, tau, lambda) are dataset-dependent and tuned, with lambda set to 0 for BBCSport and Cora.

Experimental setup

Datasets: BBCSport Caltech101-20 Cora Scene-15 MNIST-USPS ALOI-100 NoisyMNIST

Benchmarks: BBCSport Caltech101-20 Cora Scene-15 MNIST-USPS ALOI-100 NoisyMNIST

Metrics: ACC NMI PUR

Baselines: Completer MFLVC DealMVC DMCE CSPAN ADPAC SURE

Builds on: SURE Completer MFLVC granular-ball computing

Closest papers by meaning

• Contrastive Multi-view Subspace Clustering via Tensor Transformers Autoencoder (AAAI 2025 · similarity 0.88)
• EASEMVC:Efficient Dual Selection Mechanism for Deep Multi-View Clustering (CVPR 2025 · similarity 0.88)
• Medusa: A Multi-Scale High-order Contrastive Dual-Diffusion Approach for Multi-View Clustering (CVPR 2025 · similarity 0.88)
• Decoupled Contrastive Multi-View Clustering with High-Order Random Walks (AAAI 2024 · similarity 0.88)
• PROTOCOL: Partial Optimal Transport-enhanced Contrastive Learning for Imbalanced Multi-view Clustering (ICML 2025 · similarity 0.87)
• Robust Contrastive Multi-view Clustering against Dual Noisy Correspondence (NeurIPS 2024 · similarity 0.86)