Incomplete multi-view clustering by simultaneously learning robust representations and optimal graph structures

Abstract

Incomplete multi-view clustering aims to assign data samples into cohesive groups with partially available information from multiple views. In this paper, we propose a novel generalized incomplete multi-view clustering method that integrates latent representation learning, spectral embedding as well as optimal graph clustering into a unified framework. Specifically, our method first learns a set of view-specific latent representations via low-rank subspace recovery. To bridge the gap brought by the missing samples from each view, we then unify the dimensions of the latent representations and obtain the spectral embeddings by preserving the local geometric structures. Moreover, our model seeks a consensus similarity matrix with optimal cluster structures by simultaneously approximating the affinity graphs reconstructed from multiple views. As a result, both intrinsic subspace structures and the global cluster structure could be mutually optimized by exploiting the consistent information shared among view-specific available samples. Compared with seven competing techniques, the experimental results confirm the superiority of our method on two different types of incomplete multi-view clustering tasks.