Dual constraint based semi-supervised nonnegative matrix factorization for multi-view clustering

Semi-supervised nonnegative matrix factorization (NMF) has attracted considerable attentions in multi-view clustering applications. However, existing semi-supervised methods only adopt either pointwise (i.e., label) or pairwise constraints as supervisory information, without considering taking full advantage of both to further enhance the effectiveness of clustering performance. To this end, a novel dual constraint based semi-supervised nonnegative matrix factorization (DSNMF) method is proposed in this paper for multi-view clustering tasks. Concretely, a new multi-view based dual constraint (MDC) algorithm is developed in DSNMF, which simultaneously utilizes both the pointwise and pairwise supervisory information to promote the performance of multi-view clustering. Specifically, when the limited label information is obtained, the MDC algorithm not only constructs the label regularization to guide the learning of the indicator matrices, but also adopts the hypergraph based pairwise constraint propagation algorithm to construct the graph regularization. Moreover, an alternating multiplicative iterative method is developed for solving the optimization problem of DSNMF, as well as analyzing its convergence, supervisory information effect and computational complexity. Finally, numerous experimental results over five multi-view datasets conclude that DSNMF has better performance than several state-of-the-art semi-supervised multi-view clustering methods.