Multi-View Subspace Clustering With Block Diagonal Representation

Jipeng Guo, Wenbin Yin, Yanfeng Sun, Yongli Hu

Abstract

This paper introduces MSCBDR, a novel multi-view subspace clustering model designed to enforce a strict "block diagonal" structure on data representations. By fusing information from multiple complementary sources (views) and ensuring that data points only represent others within the same cluster, the model achieves superior clustering performance. Furthermore, the authors provide a mathematically proven path to convergence for its non-convex optimization, ensuring stability and reliability in practical applications.

Introduction

In the field of unsupervised learning, subspace clustering operates on the assumption that data points belonging to the same category lie within the same low-dimensional subspace. However, real-world data is increasingly "multi-view," where an object might be characterized by diverse features such as color, texture, and shape.

The primary challenge in multi-view learning is effective fusion: how to combine these disparate views into a single, accurate affinity graph. Existing methods often rely on crude fusion techniques, such as simple weighted averages, which fail to capture complex, complementary relationships. Moreover, many models struggle to ensure the resulting representation is "clean." Ideally, the representation matrix should be block diagonal, meaning points from one class should only be represented by other points in the same class, with zero correlation to other clusters.

Methodology

The proposed Multi-View Subspace Clustering with Block Diagonal Representation (MSCBDR) addresses these challenges through three core components:

  1. Self-Representation: Each data point is modeled as a linear combination of other points in the dataset ($X = XZ$).
  2. Block Diagonal Regularization: The model incorporates a specific mathematical constraint ($\|Z\|_k$) that forces the representation matrix $Z$ to have exactly $k$ connected components, directly promoting the ideal block diagonal structure.
  3. Complementary Fusion: Instead of simple averaging, the model utilizes a regularizer that encourages the global representation $Z$ to incorporate unique, discriminative information from each individual view ($Z^v$).

Because the resulting objective function is non-convex, the authors developed an alternating minimization algorithm. This approach decomposes the complex problem into four simpler sub-problems solved iteratively. Crucially, the paper provides a formal proof that this algorithm is guaranteed to converge to a stationary point.

Experimental Evaluation

The MSCBDR model was evaluated on five diverse public benchmarks, including facial recognition and object classification tasks. The datasets used include ORL (400 facial images), Yale (165 facial images under varying conditions), MSRCv1 (240 object images), COIL20 (1,440 object images), and Still DB (467 human action images).

Sample facial images from the datasets used to evaluate MSCBDR, showing variations in lighting and expression.

Key Results and Insights

MSCBDR consistently outperformed state-of-the-art baselines, such as CRSC and LMSC, across nearly all metrics:

  • Accuracy (ACC): On the Yale dataset, MSCBDR achieved 72.91% accuracy, significantly higher than the next best baseline (LMSC at 68.82%).
  • Purity: On the COIL20 dataset, the model reached a near-perfect purity of 98.68%.
  • Efficiency: Despite the mathematical complexity, the algorithm typically converges in fewer than 20 iterations across all tested datasets.

The Block Diagonal Advantage

While most subspace clustering methods aim for sparsity or low-rank structures, this research argues that the block diagonal property is a more direct and effective way to capture intrinsic cluster structure. It explicitly prevents "leakage" between different classes. Furthermore, the model prioritizes complementarity over similarity; if one view (e.g., color) is noisy, another view (e.g., shape) can "fill in the gaps" to maintain the integrity of the block diagonal structure.

Conclusion and Future Work

While MSCBDR shows exceptional performance, the authors noted limitations on datasets like MSRCv1, where high inter-class similarity makes enforcing a strict block diagonal structure difficult. Additionally, the algorithm involves an $O(n^3)$ term due to eigenvalue decompositions, which may pose scalability challenges for very large datasets.

Future research directions include adapting the model for "incomplete" views (where some data points are missing features) and optimizing the computational complexity for large-scale "Big Data" applications.

Summary updated

Abstract

Self-representation model has made good progress for a single view subspace clustering. This paper proposed the multi-view subspace clustering model based on self-representation. This model assumes that the samples from different classes are embedded in independent subspaces. Thus, the fused multi-view self-representation feature should be block diagonal, and a block diagonal regularizer with the complementarity of multi-view information is given. The model optimization algorithm by alternating minimization is proposed and its convergence without any additional assumption is proved. With the complementarity of multi-view information and the block diagonal property, our model will depict data more comprehensively than single view independently. The extensive experiments on public datasets demonstrate the effectiveness of our proposed model.
BibTeX
@article{guo2019multi,
  author = {Jipeng Guo and Wenbin Yin and Yanfeng Sun and Yongli Hu},
  title = {{Multi-View Subspace Clustering With Block Diagonal Representation}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2923614},
}
Figure