Joint learning of latent subspace and structured graph for multi-view clustering

Most existing multi-view clustering methods rely solely on subspace clustering or graph-based clustering. Subspace clustering reduces the redundant information in high dimensional data, but it neglects the intrinsic structural dependencies among samples. Graph-based clustering can model the similarity among samples but tends to suffer from redundant information. In this paper, a novel framework jointing subspace learning and structured graph learning for multi-view clustering (SSMC) is proposed, which benefits from the merits of both subspace learning and structured graph learning. SSMC utilizes graph regularized subspace learning to obtain low dimensional consensus features, where the embedded features are ensured to have maximized correlation to reduce the redundant information, and the graph regularization forces embedded features to preserve their sample similarities. Meanwhile, an adaptive structured graph is learned based on the consensus features in the embedded feature space, avoiding the curse of dimensionality in the graph learning procedure. A rank constraint forces the learned graph to have exactly the same number of connected components as the number of clusters, to obtain a more reliable structured graph. Moreover, an effective algorithm is proposed to optimize the SSMC, where the graph regularized subspace learning part and the structured graph learning part are jointly optimized in a mutual reinforcement manner. The experimental results on real-world benchmark datasets show that the SSMC outperforms the state-of-the-arts in multi-view clustering tasks.