Abstract: Binary multi-view clustering has attracted intense attention from researchers due to its efficiency in handling large-scale datasets. However, previous clustering approaches suffer from at least two limitations. First, they ignore correlations among the features of original data. As a result, the geometric consistency of data is not preserved in the to-be-learnt binary representation space. Second, redundant and noisy features mixed in original data inevitably limit the ultimate clustering performance. In light of this, we propose a novel discriminative binary multi-view clustering (DBMVC) method to address the issues. Specifically, the proposed DBMVC first maps original data onto the Hamming space to obtain corresponding binary codes, which can effectively reduce the computational complexity and storage costs in the following steps. To enable our method to select useful features from original data and get a discriminative representation, the L_2,\,1-norm is used to constrain the feature projection matrix. In addition, a graph regularization term is further introduced to preserve the local manifold structure of the learned binary representation. Finally, an alternative iterative optimization algorithm is designed to solve the optimization problems of the objective function. Comprehensive experiments on six large-scale multi-view datasets validate that the proposed DBMVC markedly outperforms other state-of-the-art methods in terms of effectiveness and efficiency.