Fine-grained multi-view clustering with robust multi-prototypes representation

Multi-view clustering is a hot research topic that improves clustering performance by leveraging complementary information from multiple views. Recently, many multi-view clustering methods have been proposed. Most of them take the entire sample space as a fusion object and treat the local structures within each view equally. This paradigm is considered coarse-grained information fusion. However, in many real-world applications, different local structures with strong or weak clustering capacities could coexist in one view. To fully exploit valuable information of local structures, it is necessary to distinguish these local structures with different clustering capacities. In this paper, we propose a novel fine-grained multi-view clustering method. First, the sample space of each view is divided into many sub-clusters by using multi-prototypes representation. Second, the robustness of the multi-prototypes representation is enhanced by reducing the overlap between sub-clusters, which can reduce the effect of noise data. Finally, each sub-cluster's contribution weights are automatically assigned based on its clustering capacity. In addition, the robust multi-prototypes representation, the fine-grained multi-view fusion, and the clustering process are integrated into a unified framework. An effective alternating optimization algorithm is adopted to solve the objective function. Extensive experiments on two toy datasets and several real-world datasets prove that our method outperforms the traditional methods in clustering accuracy.

Automated summary

Multi-view clustering is a hot research topic that leverages complementary information from multiple views to improve clustering performance. In this paper, a novel fine-grained multi-view clustering method is proposed. It divides the sample space of each view into sub-clusters using multi-prototypes representation, enhances the robustness of representation by reducing sub-cluster overlap, and assigns contribution weights based on clustering capacity. The method integrates robust multi-prototypes representation, fine-grained multi-view fusion, and clustering process into a unified framework, and achieves better clustering accuracy compared to traditional methods.