Multi-view spectral clustering by simultaneous consensus graph learning and discretization

Multi-view spectral clustering has drawn much attention due to its excellent performance in grouping arbitrarily shaped data. Most of the multi-view spectral clustering methods perform clustering, relying on multiple predefined similarity graphs. Unfortunately, they require three separate steps in sequence, i.e., similarity graph learning, cluster label relaxing, and discretization of continuous labels, resulting in a compromised clustering performance. The reason is that the predefined similarity graph may not be optimal for the subsequent clustering, and the relax-and-discretize strategy may cause significant information loss. To this end, in this work, we disentangle the above issue by simultaneous consensus graph learning and discretization, where the similarity graph and the discrete cluster label matrix are learned in a unified framework. Specifically, the consensus graph shared by all views is adaptively learned with the guidance of the discrete cluster label matrix. In contrast, the cluster information hidden in the discrete label matrix can effectively boost the quality of the consensus graph. As a result, information loss among independent steps is effectively obviated, and better performance can be achieved. Experimental results on several challenging data sets validated the effectiveness of the proposed method compared to the state-of-the-art approaches. (C) 2021 Published by Elsevier B.V.