Simultaneous multi-graph learning and clustering for multiview data

Abstract

As many data in practical applications occur or can be arranged in multiview forms, multiview clustering utilizing certain complementary and heterogeneous information in various views to promote the clustering performance, has received much attention recently. Among varieties of methods, graph-based unsupervised learning methods are an essential approach for learning intrinsic structure relations of multiview data for clustering. Most of them firstly integrate information from each view into a consensus graph, which is then fed into the classic spectral clustering to achieve clustering. Such a two-step clustering paradigm is difficult to obtain the optimal clustering results even though every step performs individual optimization. This paper integrates multi-graph construction, consensus graph construction, and clustering in a unified learning framework, which can simultaneously consider the consistency and complementarity of multiview data to provide the clustering results directly. Moreover, we treat each view differently by automatic weight learning. Specifically, multi-graph learning, consensus graph learning, and weight learning are seamlessly integrated so that the related variables can be iteratively updated in the unified optimization framework–the clustering results towards an overall optimum. Comprehensive experiments on real multiview datasets verify the superiority of the proposed method over other state-of-the-art baselines in terms of three clustering evaluation metrics.