ASFS: A novel streaming feature selection for multi-label data based on neighborhood rough set

Neighborhood rough set based online streaming feature selection methods have aroused wide concern in recent years and played a vital role in processing high-dimensional data. However, most of the existing methods are directly applied to handle single-label data, or to handle multi-label data by converting multi-label data into a combination of multiple single-label datasets, which ignores that the label set of multi-label data is an integral whole. In this paper, we propose a novel online streaming feature selection for multi-label learning via the neighborhoorough set model, in which feature significance, feature redundancy, and label space integrity are taken into account, simultaneously. To be specific, we first define a new adaptive neighborhood relation to avoid the setting of neighborhood parameter and restructure the neighborhood rough set model to be suitable for processing multi-label data directly. Based on this model, we introduce a evaluation criterion to select features that are important relative to label set and the currently selected features, and present an optimization objective function to update the selected feature subset and filter out redundant features. Comparative experiments on different types of data sets explicitly verify the advantages of the proposed method.