Unsupervised Multi-view Clustering by Squeezing Hybrid Knowledge from Cross View and Each View

Junpeng Tan, Yukai Shi, Zhijing Yang, Caizhen Wen, Liang Lin

Abstract

Multi-view clustering methods have been a focus in recent years because of their superiority in clustering performance. However, typical traditional multi-view clustering algorithms still have shortcomings in some aspects, such as removal of redundant information, utilization of various views and fusion of multi-view features. In view of these problems, this paper proposes a new multi-view clustering method, low-rank subspace multi-view clustering based on adaptive graph regularization. We construct two new data matrix decomposition models into a unified optimization model. In this framework, we address the significance of the common knowledge shared by the cross view and the unique knowledge of each view by presenting new low-rank and sparse constraints on the sparse subspace matrix. To ensure that we achieve effective sparse representation and clustering performance on the original data matrix, adaptive graph regularization and unsupervised clustering constraints are also incorporated in the proposed model to preserve the internal structural features of the data. Finally, the proposed method is compared with several state-of-the-art algorithms. Experimental results for five widely used multi-view benchmarks show that our proposed algorithm surpasses other state-of-the-art methods by a clear margin.

Introduction

In machine learning, "multi-view" data refers to information collected from different sources or perspectives—such as a video represented by pixels, audio, and text. While multiple views should theoretically improve clustering, they often introduce redundancy. Traditional multi-view clustering (MVC) methods struggle to balance consensus (finding common ground across views) and complementarity (capturing unique information in specific views). Furthermore, standard mathematical tools like nuclear norms often "over-penalize" data, leading to the loss of vital structural details.

Methodology: The UMC-CEV Framework

The proposed Unsupervised Multi-view Clustering by Squeezing Hybrid Knowledge (UMC-CEV) framework integrates two specialized sub-modules to address these challenges:

  • ML0-LSSC (Multi-view $L_0$-Motivated Subspace Clustering): This module focuses on unique information. It utilizes a novel threshold function (GMC-LRSSC) as a precise alternative to standard sparsity constraints, extracting view-specific features without over-smoothing.
  • MLRR-AGR (Multi-view Low-Rank Representation with Adaptive Graph Regularization): This module extracts common information. It employs Adaptive Graph Regularization (AGR), where the similarity graph evolves during learning to better reflect the data's internal manifold structure.
The UMC-CEV framework extracts common and unique information using adaptive graph regularization and collaboration training.

By moving beyond the nuclear norm and using a "firm threshold function," the model avoids biasing large singular values. The entire system is optimized using the Alternating Direction Method of Multipliers (ADMM) to ensure efficient convergence.

Experimental Setup

The researchers evaluated UMC-CEV against state-of-the-art methods using five diverse real-world datasets, ranging from facial images to news articles.

Summary of the five multi-view benchmarks used for evaluation.

Key Results and Discussion

UMC-CEV consistently outperformed high-profile competitors such as DiMSC, LMSC, and ECRMSC across all benchmarks:

  • Accuracy (ACC) Gains: On the ORL_mtv dataset, UMC-CEV achieved an accuracy of 91.25%, representing a ~6% improvement over the next best method.
  • Stability: The model demonstrated high stability on the uci_digit dataset, with minimal variance (1.15%) across different parameter settings.
  • Visual Clarity: Similarity matrix visualizations revealed cleaner, sharper "blocks" (clusters) in UMC-CEV compared to other methods, indicating a superior signal-to-noise ratio.

Limitations and Future Work

Despite its performance, the algorithm's $O(n^3)$ scaling may pose a bottleneck for massive datasets. Additionally, the model remains sensitive to penalty factors ($\lambda_1, \lambda_2, \lambda_3$) in certain contexts, such as text data. Future research will focus on integrating these "hybrid knowledge" insights into deep neural networks to improve scalability.

Conclusion

UMC-CEV introduces a robust strategy for multi-view clustering by simultaneously "squeezing" common and unique knowledge. By combining adaptive graph learning with flexible low-rank constraints, it sets a new performance standard for unsupervised multi-view learning tasks.

Summary updated

Abstract

Multi-view clustering methods have been a focus in recent years because of their superiority in clustering performance. However, typical traditional multi-view clustering algorithms still have shortcomings in some aspects, such as removal of redundant information, utilization of various views and fusion of multi-view features. In view of these problems, this paper proposes a new multi-view clustering method, low-rank subspace multi-view clustering based on adaptive graph regularization. We construct two new data matrix decomposition models into a unified optimization model. In this framework, we address the significance of the common knowledge shared by the cross view and the unique knowledge of each view by presenting new low-rank and sparse constraints on the sparse subspace matrix. To ensure that we achieve effective sparse representation and clustering performance on the original data matrix, adaptive graph regularization and unsupervised clustering constraints are also incorporated in the proposed model to preserve the internal structural features of the data. Finally, the proposed method is compared with several state-of-the-art algorithms. Experimental results for five widely used multi-view benchmarks show that our proposed algorithm surpasses other state-of-the-art methods by a clear margin.
BibTeX
@article{tan2021unsupervised,
  author = {Junpeng Tan and Yukai Shi and Zhijing Yang and Caizhen Wen and Liang Lin},
  title = {{Unsupervised Multi-view Clustering by Squeezing Hybrid Knowledge from Cross View and Each View}},
  journal = {IEEE Trans. Multimedia},
  year = {2021},
  doi = {10.1109/tmm.2020.3019683},
}
Figure