Binary multi-view sparse subspace clustering

Multi-view subspace clustering, which partitions multi-view data into their respective underlying subspaces, has achieved the remarkable clustering performance by extracting abundant complementary information from data of different views. However, existing subspace clustering methods almost suffer from very heavy computational burden that restricts their capacity on computational efficiency for large-scale datasets. Recently, hashing/binary code learning has attracted intensive attentions due to fast Hamming distance computation and much less storage requirement, but existing related research does not explore underlying subspace clustering structure well that widely exists in real-world data. In order to handle the both issues, in this paper, we propose a multi-view subspace clustering method named Hashing Multi-view Sparse Subspace Learning (HMSSL). HMSSL incorporates multi-view binary code learning and binary sparse subspace learning with a “thin” dictionary into a unified framework. HMSSL encodes multi-view real-valued features in the original space into compact common binary codes in the Hamming space for fast Hamming distance computation by multi-view binary code learning and learns the binary sparse subspace representation matrix for exploring the underlying subspace clustering structure efficiently and effectively by binary sparse subspace learning with a “thin” dictionary matrix. All the columns of the dictionary matrix are randomly and uniformly sampled from all the columns of the compact common binary code matrix. We design an effective binary optimization algorithm based on alternating direction multiplier method and analyze its time complexity. Extensive experiments performed on six benchmark multi-view datasets demonstrate the effectiveness of HMSSL in comparison with ten state-of-the-art baselines in this field.