Dual Global Structure Preservation Based Supervised Feature Selection

摘要

The recent literature indicates that the global structure preservation is very important for sparse representation based supervised feature selection. However, the selected features in preserving different global structures are often different and which global structure is the best not yet known. As a result, which feature selection result we should trust is confusing. The reason may be that each global structure does not carry enough information for the data, as the distribution of a real life data is very complex. To overcome the above problem, in this paper, a dual global structure preservation based supervised feature selection (DGSPSFS) method is proposed. In DGSPSFS, the supervised dimensional reduction method based on manifold learning is used to calculate the response matrix, which can contain more information of the data. And a new sparse representation framework that can preserve two global structures in the same time is proposed, which can comprehensively use two response matrices to fully utilize the information of the data. As a result, the features that can carry more information are selected. A comprehensive experimental study is then conducted in order to compare our feature selection algorithms with many state-of-the art ones in supervised learning scenarios. The conducted experiments validate the effectiveness of our feature selection.