Auto-weighted multi-view clustering via spectral embedding

Abstract

As is well-known, multi-view clustering has attracted considerable attention since many benchmark data sets exist heterogeneous features. Previous multi-view spectral clustering methods mainly contain two steps: 1) constructing multiple similarity graphs; 2) performing K-means (KM) clustering. The two-step strategy cannot acquire optimal results since the clustering performance highly relies on the constructed similarity graphs. To address this drawback, a unified framework named as an Auto-weighted Multi-view Clustering via Spectral Embedding (AMCSE) is presented. In the new proposed method, it can consider the clustering capacity heterogeneity of different views as well as directly obtain the clustering results. More importantly, the unified framework can make multiple graph learning guide the clustering result discretization, while the latter is in turn to conduct to learn better spectral embedding. A series of experiments are conducted on six real-world data sets, and the clustering results verify that the proposed method is not only effective but also efficient, comparing with state-of-the-art graph-based multi-view clustering approaches.