Multi-view Spectral Clustering with High-order Similarity Learning

Abstract

AbstractClustering objects with diverse attributes obtained from multiple views is full of challenges in fusing the multi-view information. Many of the present multi-view clustering (MVC) methods concentrate on direct similarity learning among data points and fail to excavate the hidden high-order similarity among different views. Therefore, it is difficult to obtain a dependable clustering assignments. To address this problem, we propose the high-order similarity (HOS) learning model for multi-view spectral clustering (MCHSL). The proposed MCHSL learns the first-order similarity (FOS), second-order similarity (SOS), and the HOS collaboratively to excavate the local structure relations, proximity structure relations of paired data points and the interactive-view relations among different views instead of the common similarity learning. Then spectral clustering is performed to obtain the final clustering assignments. Extensive experiments performed on some public datasets indicate that the proposed MCHSL has better clustering performance than benchmark methods in most cases and is able to reveal a dependable underlying similarity structure hidden in multiple views.KeywordsFirst-order similaritySecond-order similarityHigh-order similarityMulti-view clustering