Cross-View Diversity Embedded Consensus Learning for Multi-View Clustering

Chong Peng, Kai Zhang, Yongyong Chen, Chenglizhao Chen, Qiang Cheng

Abstract

Multi-view clustering (MVC) is essential for analyzing data from multiple sources, such as video, audio, and text. However, traditional MVC methods often fail to capture high-order relationships between data points and struggle to balance global consensus with view-specific diversity. This paper introduces CCL-MVC (Cross-View Diversity Embedded Consensus Learning), a novel framework that addresses these limitations using cross-order neighbor tensors and non-convex optimization. Experimental results across six benchmark datasets demonstrate that CCL-MVC significantly outperforms 17 state-of-the-art baselines, particularly on complex datasets like StillDB.

Introduction

In real-world applications, data is frequently multi-modal. For instance, a news story may consist of text, images, and video. Multi-view clustering aims to leverage the complementary information across these perspectives to improve grouping accuracy. Despite its potential, MVC faces three primary hurdles:

  1. Inaccurate Approximations: Reliance on convex approximations like the Tensor Nuclear Norm can lead to imprecise results.
  2. Neglect of High-Order Relationships: Most methods focus on immediate (first-order) neighbors, missing deeper latent connections.
  3. Consensus-Diversity Imbalance: It is challenging to find a consensus representation that doesn't discard the unique, diverse information inherent in individual views.

Methodology

The proposed CCL-MVC framework introduces several innovations to overcome these challenges:

Cross-Order Neighbor Tensor

Instead of relying on standard graphs, CCL-MVC constructs a tensor based on $K$-th order Markov transition probabilities. This effectively simulates multi-step random walks, uncovering deeper structural relationships—essentially finding "friends of friends" within the data.

Non-Convex Learning

To improve upon standard convex methods, the authors utilize Log-based Tensor Rank Approximation (LTRA) and Log-based Tensor Sparsity Approximation (LTSA). These functions more accurately reflect the true rank and sparsity of the data compared to traditional nuclear norms.

Diversity Embedding

CCL-MVC explicitly incorporates a diversity term into its objective function. This prevents the model from over-weighting redundant views, forcing the consensus matrix to capture the most unique and discriminative signals from every available perspective.

Efficient Optimization

To handle the computational complexity of tensor updates, the authors derived a closed-form solution using the Sherman-Morrison-Woodbury formula. This approach avoids the $O(n^3V^3)$ complexity typically associated with such models, ensuring scalability and speed.

Experimental Results

The model was evaluated on six benchmark datasets, including BBC news articles, flower images, handwritten digits (UCI-3view), and complex scene datasets (StillDB, MITindoor).

Performance Comparison

CCL-MVC achieved the best performance across all 24 test cases (6 datasets $\times$ 4 metrics). Notably, on the "difficult" StillDB dataset, CCL-MVC improved clustering accuracy by 14% to 17% over the next best performing methods.

Performance comparison of CCL-MVC against 17 state-of-the-art methods across six datasets.

Impact of High-Order Relationships

Ablation studies revealed that increasing the neighbor depth from first-order to third-order ($K=3$) significantly boosts performance. However, beyond $K=3$, performance tends to plateau or decline due to the introduction of noise.

Performance analysis across different neighbor orders (K) for BBC-Sport and StillDB datasets.

Discussion and Conclusion

CCL-MVC demonstrates that incorporating high-order relationships and diversity constraints leads to more robust and accurate multi-view clustering. The use of non-convex approximations and efficient closed-form optimization makes it both precise and practical.

While the model is highly effective, future work will focus on automating the selection of the optimal order $K$ and exploring anchor-based methods to further enhance scalability for massive datasets. For practical applications, the authors recommend a third-order ($K=3$) configuration as a robust starting point.

Summary updated

BibTeX
@inproceedings{peng2024cross,
  author = {Chong Peng and Kai Zhang and Yongyong Chen and Chenglizhao Chen and Qiang Cheng},
  title = {{Cross-View Diversity Embedded Consensus Learning for Multi-View Clustering}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure