Learning missing instances in latent space for incomplete multi-view clustering

摘要

Real world objects can usually be described from multiple views. How to combine multiple views to make better use of these data has attracted a lot of attentions, and numerous multi-view clustering algorithms have been proposed. However, the data in real scenes are often incomplete. We propose an incomplete multi-view clustering method (MISS), which simultaneously learns Missing Instances and Self-representations in latent Space. Instead of filling missing instances in the original data space, we learn missing instances and self-representations in the latent space so that we can grasp the features of missing instances, while preserving the original latent spatial structure of the data. In addition, a novel fusion matrix is introduced. The missing instances are represented by linear combinations of the known instances in the latent space. This fusion matrix will be learned simultaneously according to the learning of self-representation in the latent space for multi-view clustering. Furthermore, a completeness constraint is proposed to guarantee that learning direction of missing instances is close to the original data as possible. Because of the focus on the features of missing instances and the keeping of the original data structure, MISS can not only handle randomly data missing, where some instances are randomly removed from each view, but also continuous data missing, i.e., all instances of some clusters are missing completely. Experimental results on seven datasets show that MISS achieves improvements over the state-of-the-art incomplete multi-view clustering methods.