No abstract

No abstract

Many problems in computer vision involve data sets with multiple views where observations are represented by multiple sources of features. To integrate information from multiple views in the unsupervised setting, multi-view clustering algorithms have been developed to cluster multiple views simultaneously. However, most of these algorithms only consider the situation that each example appears in all views, but can not deal with the case that the example may miss partial views data, which is often occurred in real applications, i.e., when a robot is investigating, the data may not be captured completely due to certain sensor fault. In this paper, we present a nonnegative matrix factorization method for partial multi-view clustering, which incorporates the cluster similarity and manifold preserving constraints into the unified framework. The basic principle of our proposed double constrained NMF (DCNMF), is pushing clustering solution of the same example in different views towards a common membership matrix, and meanwhile keeping the latent geometric structure of instances in the same view. Moreover, we develop the corresponding optimization scheme for our proposed method. Experiments on two-view datasets demonstrate the advantages of our proposed approach.

No abstract