Accelerated Semi-supervised Feature Selection via Adaptive Bipartite Graph

Recent years have witnessed the proliferation of semi-supervised feature selection, which can select a subset of discrimi-native features for subsequent tasks with a small amount of class label information. However, traditional methods cannot efficiently handle large-scale problems and often fail to mine the reliable similarity structure of data. To address these issues, a novel model is proposed in this paper, called Accelerated Semi-supervised Feature Selection (ASFS). Specifically, a bipartite graph between samples and anchors is adaptively constructed in the feature projection subspace to significantly reduce the computational costs of graph learning and solution procedures, so that the main computational complexity of ASFS is linearly dependent on the number of training data. Moreover, graph learning and feature selection are integrated into a unified framework, wherein they can benefit from each other. Therefore, the interference of noisy features can be largely alleviated, and meanwhile, more informative features will be selected under the guidance of the learned similarity graph. The effectiveness and efficiency of ASFS are validated by extensive experiments on multiple datasets.