Abstract: LSPE (Locality and similarity preserving embedding) feature selection algorithm firstly maintains the locality of the data based on the pre-defined graph structure of the KNN,and then maintains the locality and similarity of the data based on the low-dimensional reconstruction coefficients that define the learning data of the graph.The two steps are independent and lack of interaction.Since the number of nearest neighbors is artificially defined,the learned graph structure does not have adaptive nearest neighbors and is not optimal,which will affect the performance of the algorithm.In order to optimize the performance of LSPE,an improved locality and similarity preserving feature selection algorithm is proposed.The proposed algorithm incorporates graph learning,sparse reconstruction and feature selection into the same framework,making graph learning and sparse coding are carried out simultaneously.The coding process is required to to be sparse,adaptive neighbor and non-negative.The goal is to find a proJection that can maintain the locality and similarity of the data,and apply a l_2,1-norm to the proJection matrix,and then select the relevant features that can maintain locality and similarity.Experimental results show that the improved algorithm reduces the subJective influence,eliminates the instability of selecting features,is more robust to data noise,and improves the accuracy of ima-ge classification.

Key words: Feature selection, Locality and similarity preserving, Sparse reconstruction, Unsupervised learning