基于自适应图学习的半监督特征选择

With the increasing feature dimensionality, how to select a relevant feature subset in the case of a few labeled and large amount of unlabeled high-dimensional samples has become a hot issue in feature selection. However, existing semi-supervised feature selection algorithms directly ignore the interaction between feature selection and local structure learning, making it difficult to obtain the distribution structure information. To these ends, a semi-supervised feature selection algorithm with adaptive graph learning(SFSAG) is developed in this paper. Firstly, the label propagation is used to link the tasks of sparse projection learning on the original feature space and construction of affinity graph, such that the feature selection and local structure learning can be performed simultaneously. Then, a reliable neighbor graph is adaptively constructed by using the similarity information of samples in the projected feature space, which largely alleviates the adverse effects of noisy dimensions and facilitates selecting more discriminative features. Extensive experiments are conducted on various datasets, and the results demonstrate the effectiveness of the proposed SFSAG and its superiority in comparison with the state-of-the-art feature selection algorithms.