Abstract
Multi-view clustering (MVC) aims to integrate diverse data sources to uncover underlying structures in complex datasets. However, existing methods often suffer from biased convex approximations, neglect high-order relationships, and struggle to balance consensus with diversity. We propose Cross-View Diversity Embedded Consensus Learning (CCL-MVC), a framework that utilizes Markov random walks to capture high-order structural information and non-convex log-based tensor rank approximations for precise essential tensor recovery. By explicitly embedding cross-view diversity, CCL-MVC exploits unique discriminative information across views. Extensive experiments on six benchmark datasets demonstrate that CCL-MVC consistently outperforms 17 state-of-the-art methods, particularly on challenging datasets like StillDB.
Introduction
In the era of big data, objects are often characterized by multiple views (e.g., image pixels, text tags, and metadata). Multi-View Clustering (MVC) seeks to group these objects by integrating these diverse sources. Despite its potential, current MVC approaches face three primary hurdles: inaccurate convex approximations (like the Tensor Nuclear Norm), a focus on immediate first-order neighbors that misses latent connections, and the difficulty of finding a consensus representation that respects the unique diversity of each view.
Methodology
The CCL-MVC framework addresses these challenges through three core innovations:
- Cross-Order Neighbor Tensors: Instead of simple similarity graphs, the model employs Markov random walks to calculate the probability of reaching one sample from another in multiple steps. This captures high-order structural information crucial for defining clear clusters.
- Non-Convex Log-Based Learning: To recover the "essential tensor"—the noise-free core of the data—the framework replaces standard convex approximations with Log-based Tensor Rank Approximation (LTRA) and Log-based Tensor Sparsity Approximation (LTSA). These provide a tighter, more accurate estimation of the data's true rank.
- Diversity Embedding: A specific penalty term prevents the weights of redundant views from being simultaneously large, forcing the model to exploit unique discriminative information from each source.
For optimization, the authors developed an algorithm based on the Augmented Lagrange Multiplier (ALM), utilizing the Sherman-Morrison-Woodbury formula to ensure computational efficiency even with complex tensor operations.
Experimental Evaluation
The model was evaluated on six diverse benchmark datasets, including news articles (BBC-Sport), image data (Flowers), and complex scene recognition (MITindoor). It was compared against 17 state-of-the-art baselines, including recent methods like OMVFC-LICAG and E2OMVC.
Key Results
CCL-MVC demonstrated dominant performance across all metrics (ACC, NMI, AR, and F-Score):
- Superior Accuracy: The method achieved the best results in all 24 test cases across the six datasets.
- Gains on Challenging Data: On the StillDB dataset, where most baselines struggle to exceed 0.4 accuracy, CCL-MVC improved performance by 14% to 17% over the nearest competitor.
- Convergence: The optimization algorithm typically converges within 50 iterations, making it viable for practical applications.
Conclusion
CCL-MVC advances multi-view clustering by effectively capturing high-order relationships and utilizing non-convex approximations to recover essential data structures. The explicit modeling of cross-view diversity ensures a comprehensive consensus representation. Future work may focus on automated selection of the optimal neighbor order and improving scalability for massive datasets with millions of samples.
Summary updated