Double embedding-transfer-based multi-view spectral clustering

Abstract

It is crucial for multi-view clustering to mine the consistency among multiple views; this is useful for improving the performance of multi-view clustering, especially with noisy data. In this paper, a novel clustering model is proposed for learning consistency using multi-view data, namely double embedding-transfer-based multi-view spectral clustering (DETMSC). In contrast to many existing methods, DETMSC can simultaneously learn consistency embedding and feature embedding in a unified framework. This is due to the power of the implemented knowledge transfer between consistency embedding and feature embedding among multiple views, which is achieved through bipartite graph co-clustering. In this way, DETMSC can simultaneously consider both the consistency of multiple views and the diversity of the feature embedding for each view. Optimal consistency embedding may be approached using the diversity of feature embedding, which can further result in a good performance for multi-view clustering. Better consistency embedding can in turn help with learning the diversity of feature embedding. The DETMSC method is also robust to the noise within each view because it uses the sparse constraint in feature embedding. Extensive experimental results on several real-world benchmark datasets show that the proposed model outperforms other state-of-the-art multi-view clustering algorithms.