High-dimensional multi-label learning based on missing features

Multi-label learning mainly deals with the problem that each sample data is associated with multiple class labels, but it is difficult to obtained complete feature information at one time in practical applications.Existing multi-label learning approaches to solve the missing features, but they are not considered the missing features in high-dimensional environments and most of the existing feature dimensional reduction methods are either directly transformed from single-label feature selection methods, or cannot make full use of label information. Thus,they may not be able to get an optimal feature selection result shared by multiple labels.Motivated by this,we propose a missing feature multi-label learning method in a high-dimensional environment.Firstly, a new complementary feature matrix is obtained by learning the feature correlation matrix, which has richer complete feature information than the original missing feature matrix.Secondly, the information theory method is introduced to propose a general global optimization framework, which considers feature correlation, label correlation and feature redundancy, and realizes feature dimensional reduction of high-dimensional multi-label data.After that, to improve the performance of multi-label classification, we constrain feature correlation on coefficient matrix by assuming that if two features are strongly correlated,the similarity between their corresponding parameter vector will be large. we also constrain label correlation on output of labels to capture more sufficient relationships between different labels. Extensive experiments show a competitive performance of proposed method against other state-of the-art multi-label learning approaches.