Efficient multi-view semi-supervised feature selection

Multi-view semi-supervised feature selection can identify a feature subset from heterogeneous feature spaces of data. However, existing methods fail in handling large-scale data since they have to calculate the inverses of high-order dense matrices. Moreover, traditional methods often pre-construct graphs to mine the similarity structure of data, such that the interaction between graph construction and feature selection is directly ignored, degrading their effectiveness in practice. To address these issues, we propose an efficient multi-view feature selection method (EMSFS), which combines graph learning, label propagation as well as multi-view feature selection within a unified framework. Specifically, EMSFS can adaptively learn a bipartite graph between training samples and generated anchors, not only reducing the cost of graph computation but also tactfully avoiding the inverse of a high-order matrix. As a result, the main computational complexity of EMSFS is approximately linear to the number of training samples. Meanwhile, EMSFS simultaneously selects important features and exploits the similarity structure in the projected feature space, which enhances the reliability of the graph and positively facilitates feature selection. To solve the formulated objective function, we developed an alternating optimization, and experiments validate the effectiveness and the efficiency of EMSFS.