MASTER: A Multi-granularity Invariant Structure Clustering Scheme for Multi-view Clustering

Suixue Wang, Shilin Zhang, Qingchen Zhang, Peng Li, Weiliang HuoIJCAI 2025 · DOI

TL;DR

This paper introduces MASTER, a novel multi-view clustering framework that addresses the limitations of existing methods by extracting information at multiple granularities (sample, neighborhood, and category) across different feature spaces. By integrating these multi-level structures, MASTER achieves state-of-the-art performance on various benchmark datasets, offering a more robust and comprehensive approach to multi-view data analysis.

What Problem Does This Paper Address?

Multi-view clustering (MVC) aims to group data points based on information from multiple sources or "views." While deep learning has significantly advanced MVC, current methods often struggle with two key issues. Firstly, they tend to operate within a single feature space, neglecting the rich structural information present at different levels of abstraction (e.g., individual samples, their local neighborhoods, and broader categories). Secondly, existing reconstruction-based methods, while aiming for generalized representations, can dilute the discriminative power of view-specific information, leading to suboptimal clustering, especially with complex data distributions. This paper tackles these challenges by proposing a method that systematically captures and integrates structural information across multiple granularities and feature spaces.

Methods

MASTER employs a bottom-up, multi-granularity approach, comprising three core strategies:

  • Sample-oriented Representation Learning (SRL): This strategy focuses on the low-level feature space and individual samples. It uses an encoder-decoder architecture for self-learning reconstruction within each view. Crucially, it incorporates an information-theoretic overclustering mechanism that treats each sample as a single cluster. This maximizes mutual information between views, enhancing the discriminative ability of view-specific representations by capturing invariant sample structures.
  • Reliable Neighborhood-driven Diffusion Strategy (RND): Moving to the high-level feature space, RND models local consistency. It uses a feature mapping function to generate high-level representations and then employs a cross-view neighborhood distillation mechanism. This mechanism imitates data diffusion by adaptively refining neighbor relationships based on cluster assignments. It dynamically captures reliable affinity relationships, learning invariant local structures by encouraging intra-neighborhood similarity and inter-neighborhood dissimilarity.
  • Semantics-invariant Cluster Strategy (SIC): This strategy operates in the semantic space to capture global structure. It defines dual divergences (simplex and orthogonal constraints) between assignment predictions from different views. By leveraging the geometry of the semantic space, SIC aims to maximize inter-cluster separation and intra-cluster compactness, ensuring robust global structure learning.

These three components are integrated into a unified framework, with their respective losses combined to optimize the overall model.

Datasets

MASTER was evaluated on 8 real-world datasets:

  • BDGP (2 views)
  • CCV (3 views)
  • MNIST-USPS (2 views)
  • Fashion-MV (3 views)
  • Caltech-2V (2 views)
  • Caltech-3V (3 views)
  • Caltech-4V (4 views)
  • Caltech-5V (5 views)

The paper does not introduce new benchmarks but uses established datasets commonly used in multi-view clustering research.

Key Insights and Ideas

  • Multi-granularity is Key: The paper strongly advocates for the importance of considering data structure at multiple levels (sample, neighborhood, category) rather than relying on a single representation. This hierarchical approach allows for a more comprehensive understanding of data relationships.
  • Information Theory for Robustness: The use of information-theoretic overclustering in SRL helps to distill complementary and consistent information, making the representations more robust to noise and variations across views.
  • Data Diffusion as a Clustering Mechanism: RND's novel approach of mimicking data diffusion to learn neighborhood relationships is an interesting way to capture local structure dynamically, adapting to the evolving cluster assignments.
  • Geometric Constraints for Global Structure: SIC's reliance on simplex and orthogonal constraints in the semantic space provides a principled way to enforce global structure, ensuring well-separated and compact clusters.

Key Results

MASTER consistently outperformed 11 state-of-the-art comparison methods across all 8 benchmark datasets. The paper reports significant gains, particularly on datasets with more views (e.g., Caltech-5V), where MASTER achieved notable improvements in Accuracy (ACC), Normalized Mutual Information (NMI), and Purity (PUR). For instance, on Caltech-5V, MASTER outperformed the second-best method by 2.10% in ACC, 1.78% in NMI, and 2.17% in PUA. Ablation studies confirmed the effectiveness of each component (SRL, RND, SIC), showing that removing any one of them led to a decrease in performance. Hyperparameter analysis indicated that MASTER is robust to the selection of trade-off parameters.

Limitations and Future Work

The paper does not explicitly detail limitations or future work directions. However, based on the methodology, potential areas for future research could include:

  • Scalability: While the paper demonstrates strong performance, exploring the scalability of MASTER to extremely large datasets or a very high number of views might be a future direction.
  • Adaptability to Different Data Modalities: Investigating how MASTER's multi-granularity approach can be adapted to diverse data modalities beyond image and text could be beneficial.
  • Automated Hyperparameter Tuning: While the paper shows robustness, further research into more automated methods for hyperparameter tuning could enhance its practical applicability.

Summary updated

BibTeX
@inproceedings{wang2025master,
  author = {Suixue Wang and Shilin Zhang and Qingchen Zhang and Peng Li and Weiliang Huo},
  title = {{MASTER: A Multi-granularity Invariant Structure Clustering Scheme for Multi-view Clustering}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/714},
}
Figure