Multi-order similarity learning for multi-view spectral clustering

Abstract

This paper explores the problem of multi-view spectral clustering (MVSC) based on multi-order similarity learning. Unlike the existing methods that focus on direct similarity of pairwise data points without considering the hidden multi-order similarity among different data points, a novel multi-order similarity learning model for MVSC (MOSL) is proposed. Specifically, the first-order similarity (FOS) and second-order similarity (SOS) are learned to excavate the local structure relation and adjacent structure relation of pairwise data points. Afterwards, the third-order similarity (TOS) based on low-rank tensor is learned to excavate the view-specific information and consensus information from multiple views. Moreover, a trace constraint on each affinity graph from multiple views is learned to ensure the strict block diagonal structure of each affinity graph. Extensive experiments on six commonly benchmark datasets show that the proposed method outperforms state-of-the-art methods in most scenarios and is capable of revealing a reliable affinity graph structure concealed in different data points.