Unsupervised Feature Selection Based On Subspace Learning

With the rapid development of modern science and technology,data is becoming more complex,resulting in problems such as dimension explosion.To reduce data complexity,dimension reduction has become an important research field of machine learning and data mining.As an essential dimension reduction technology,feature selection has received more and more attention.In recent years,many feature selection algorithms have been proposed,but the discriminant information and the self-correlation of data have not been well utilized.To solve the problem of not utilizing the discriminative information of data and the correlation between data itself in feature selection,this paper introduces discriminant learning and self-representation learning on the unsupervised feature selection algorithm based on subspace learning.The main research contents of this paper are as follows:（1）An unsupervised feature selection algorithm based on discriminant virtual label regression is proposed to address the issue of unsupervised feature selection not utilizing discriminative information from data.The algorithm obtains a low dimensional representation of data through subspace learning.Then the relationship between label space and low dimensional subspace is explored through virtual label regression,and the learned labels are used to guide the learning of linear discriminant analysis.The feature selection matrix is constrained by thel_2,1-norm.Through a large number of experiments on different data sets,we can conclude that this method can achieve the best feature selection performance.（2）A novel unsupervised feature selection algorithm based on latent energy embedding is proposed to address the issue of unsupervised feature selection not utilizing the correlation of data itself.By rewriting subspace learning into the form of energy preservation,the main energy of data of data is fully preserved,and the reconstructed data is used as new data for self-representation learning to fully reduce noise interference.The method also maintains the global and local geometric structure of the data by adding low-rank and sparse constraints,while introducing manifold learning to make full use of the manifold information of the data.Finally,thel_2,1-norm constraint is applied to the feature selection matrix to maintain the sparsity of features.Through a large number of experiments on different data sets,we conclude that the proposed method can achieve the best feature selection performance.