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

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

TL;DR

This paper introduces UMC-CEV, a novel unsupervised framework that boosts multi-view clustering by simultaneously "squeezing" common knowledge shared across different data views and unique features found only in individual views. By combining adaptive graph learning with a more flexible low-rank constraint, the authors outperform existing state-of-the-art methods across five major benchmarks by a significant margin.

What Problem Does This Paper Address?

In the world of machine learning, "multi-view" data refers to information collected from different sources or perspectives—for example, a video can be represented by its raw pixels, its audio track, and its text captions. While having more views should lead to better clustering, it often introduces a headache: redundancy.

Traditional multi-view clustering (MVC) methods often struggle to balance two competing needs:

  1. Consensus: Finding the "common ground" that all views agree on.
  2. Complementarity: Capturing the unique, specialized information that only exists in one specific view.

Furthermore, many existing models rely on "nuclear norms" or "$L_1$-norms" to enforce sparsity and low-rank structures. The authors argue these mathematical tools are too blunt—they "over-penalize" the data, leading to approximations that lose vital structural details.

Methods

The authors propose a unified framework called Unsupervised Multi-view Clustering by Squeezing Hybrid Knowledge (UMC-CEV). It integrates two specialized sub-modules:

  • ML0-LSSC (Multi-view $L_0$-Motivated Subspace Clustering): This module focuses on the "unique" information. It uses a novel threshold function (GMC-LRSSC) that acts as a more precise alternative to standard sparsity constraints, allowing the model to extract view-specific features without over-smoothing them.
  • MLRR-AGR (Multi-view Low-Rank Representation with Adaptive Graph Regularization): This module focuses on the "common" information. Crucially, it uses Adaptive Graph Regularization (AGR). Instead of using a fixed similarity graph, the graph evolves during the learning process to better reflect the internal manifold structure of the data.

The model is optimized using the Alternating Direction Method of Multipliers (ADMM), a popular algorithm for breaking complex optimization problems into smaller, solvable steps.

Datasets

The researchers evaluated their model on five diverse real-world datasets:

  1. ORL_mtv: 400 grayscale images of faces under different conditions.
  2. Extended YaleB: 1,102 face images of 38 individuals under varying lights.
  3. 3-sources: 169 news stories collected from BBC, Reuters, and The Guardian.
  4. uci_digit: 2,000 examples of handwritten digits (0-9).
  5. BBCSport: 116 articles from five subject areas in sports news.

Key Insights and Ideas

  • The "Hybrid Knowledge" Strategy: Most models pick a side—either they look for consensus or they look for diversity. UMC-CEV argues that you need both to truly "squeeze" the knowledge out of the data.
  • Moving Beyond the Nuclear Norm: By using a "firm threshold function" instead of a standard nuclear norm, the model avoids the bias of over-penalizing large singular values, preserving more of the original data's signal.
  • Graphs That Learn: Most clustering methods assume they know how "similar" two points are at the start. By using an adaptive graph, UMC-CEV lets the data's own structure dictate similarity as the clustering improves.

Key Results

The proposed UMC-CEV method consistently outperformed several high-profile competitors (like DiMSC, LMSC, and ECRMSC).

  • Accuracy (ACC) Gains: On the ORL_mtv dataset, UMC-CEV achieved an accuracy of 91.25%, a roughly 6% improvement over the second-best method.
  • Stability: On the uci_digit dataset, the model showed remarkable stability, with the error between maximum and minimum accuracies across different parameter settings being only 1.15%.
  • Visual Clarity: The authors provided visualizations of the "Similarity Matrix." In UMC-CEV, the resulting matrices showed much cleaner, sharper "blocks" (representing clusters) compared to other methods, indicating a much higher signal-to-noise ratio.

Limitations and Future Work

While the results are impressive, the paper notes a few areas for further exploration:

  • Computational Complexity: The algorithm has a complexity of roughly $O(tn_v(n^3 + mn^2))$. While the authors argue this is "feasible in theory," $O(n^3)$ scaling can become a bottleneck for truly massive datasets with millions of samples.
  • Parameter Sensitivity: While stable on some datasets, the model is sensitive to its penalty factors ($\lambda_1, \lambda_2, \lambda_3$) on others, like the 3-sources text dataset. Finding the "sweet spot" for these parameters still requires some manual tuning.
  • Scaling to Deep Learning: The current work is a traditional matrix-decomposition approach. Integrating these "hybrid knowledge" insights into deep neural networks (Deep MVC) is a logical next step.

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