Multi-view Ensemble Clustering Based on BLS-Autoencoder

With the richness of data sources, data from various modalities reflect the features of samples from different views. To uncover the cluster structure of these data in an unsupervised setting, multi-view clustering approaches have garnered extensive attention. However, the current multi-view clustering approaches usually learn a shared data representation or graph embedding for the final clustering process, which may lead to the loss of view-specific information. Therefore, in this paper, we propose a multi-view ensemble clustering framework based on BLS-autoencoder(MVEC-BLSAE). In generation phase, a multi-view BLS-autoencoder(MVBLSAE) considering both the cross-view subspace structure and the within-view feature projection is designed, which can extract compressed feature subspaces and contribute to producing multiple basic partitions. In integration phase, based on the basic partitions, the sample-cluster relevance is captured through computing fuzzy partition matrices, which are then leveraged to achieve a better consensus partition via graph-cut algorithm on the constructed bipartite graph. According to the experimental results of parameters analysis, comparative analysis and ablation study, MVEC-BLSAE achieves more satisfactory performance compared to several existing multi-view clustering approaches.