Direct multi-view spectral clustering with consistent kernelized graph and convolved nonnegative representation

Multi-view clustering attempts to partition unlabeled objects into clusters by making full use of complementary and consistent information in the features of multiple views. Most existing methods perform this task in three sequential phases: Estimating individual or consistent similarity matrices, spectral embedding, and cluster partitioning. In this paper, we present a novel method that can overcome some of the shortcomings of previous multi-view clustering methods. Our approach is called "Multi-view Clustering via Kernelized Graph and Nonnegative Embedding". Based on a single global criterion, it can jointly provide the consistent similarity matrix for all views, the consistent spectral representation, the soft cluster assignments, and the view weights. To our knowledge, our approach is the first to combine all these unknown matrices into a single criterion. Our proposed scheme has two interesting properties that the recent works do not have simultaneously. First, the proposed approach does not require an additional clustering step since the clustering assignments are solved directly. Second, the soft cluster assignments are directly linked to the representation of the views. Several experiments on real datasets demonstrate the effectiveness of the proposed method. It performs well compared to many competing methods.