Multi-view Clustering Based on Collaborative Reconstruction

Multi-view clustering is an important method to mine effective information from multi-view data in unsupervised learning environment. Due to the complex relationship between different data views, high-performance learning algorithms for multi-view data should not only consider the consistency and complementarity between different views simultaneously but also reduce the impact of adverse private information of single view. In this paper, we propose a novel multi-view clustering network, called Multi-view Clustering based on Collaborative Reconstruction (MCCR), to address these issues. This method captures consistency and complementarity information from different views in a fusion-free manner. Specifically, each view is pre-trained with an AutoEncoder to extract low-dimensional features from the original space. Subsequently, a main view is selected, whose latent representation features are imported into the decoders of each view for collaborative reconstruction. This strategy guides each autoencoder to focus on not only the extraction of private information, but also the consistent information of multiple views. At the same time, the encoder of the main view can also obtain the complementary information of other views after back-propagation. Finally, a self-expression layer without bias is used to learn the affinity matrix for the latent representation, and the clustering result is obtained by applying spectral clustering to the learned affinity matrix. Extensive experiments on public data sets have verified the advantage of proposed method.