Multi-view co-clustering with multi-similarity

Multi-view co-clustering, which clustering the two dimensions of samples and features of multi-view data at the same time, has attracted much attention in recent years. It aims to exploit the duality of multi-view data to get better clustering results. However, most of the existing multi-view co-clustering algorithms consider the sample-feature information of the data while ignoring the sample-sample, feature-feature information, and thus cannot fully mine the potential information contained in the data. Therefore, this paper proposes a multi-view co-clustering based on multi-similarity. In particular, based on spectral clustering, we propose a method of constructing graph to improve the performance of clustering, which is no longer limited to the relevance between samples and features. At the same time, inspired by the ensemble algorithm, we use multiple co-clustering algorithms to calculate the similarity information of each view data, which makes the algorithm more robust. Compared with the existing multi-view co-clustering methods, the proposed algorithm exploits the more comprehensive similarity information in each view data, including sample-sample, feature-feature, and sample-feature similarity information. We performed experiments on several benchmark datasets. Due to mining and using more similarity information, our experimental results are better than the comparison method in the three evaluation indicators. In particular, on some data with co-occurrence features such as (word-document), our algorithm achieves better results and can obtain higher accuracy.