Deep Multi-view Sparse Subspace Clustering

Abstract

Most multi-view subspace clustering algorithms construct the affinity matrix with shallow features extracted from each view separately. The integration of multi-view features are left for extended spectral clustering algorithm. The lack of deep feature extraction and interaction across different views prevents the effective exploration of information complementary for multi-view datasets. To address this problem, this paper proposes a novel deep multi-view sparse subspace clustering (DMVSSC) model which consists of convolutional auto-encoders (CAEs) and CCA-based self-expressive module. The proposed model can not only extract deep features of each view data with few parameters but also integrate multi-view features based on CCA. Furthermore, a two-stage joint optimization strategy is proposed for tuning the whole model. Experiments on four benchmark data sets show that our proposed model significantly outperforms the state-of-the-art multi-view subspace clustering algorithms.