Unsupervised Feature Selection Based On Graph Regularization

With the rapid development of computer technology,the information contained in data becomes more and more complicated,which leads to the explosive growth of data dimension.However,there only a small part of information is truly useful in the original high dimensional data.Without proper processing,these high dimensional data would lead to the burden of storage and usage.Thus,the dimensionality reduction methods have aroused widespread attention.As an efficient dimensionality reduction method,feature selection has become a popular research topic in fields of machine learning and data mining.In recent years,various of unsupervised feature selection algorithms have been proposed.But,the local information and discriminant information which hidden in data are easily overlooked.To cope with these two issues,this thesis proposes the method as follows:(1)An unsupervised feature selection method called dual-graph regularized subspace learning based feature selection(DGSLFS)is proposed.Firstly,DGSLFS uses the framework of subspace learning based feature selection to deal with high dimensional data and find a proper low dimensional data space representation for it.Then,to fully explore the useful information that hidden in the local geometric structure of data,DGSLFS adds the graph regurlarization term in both sample space and data space.Finally,the loss function and feature selection matrix is constrained by the-norm,which can promote the robustness of method and ensure the sparsity of features.(2)An unsupervised feature selection method called graph regularized virtual label regression(GVLR)for unsupervised feature selection is proposed.To deal the high dimensional data in a better way,GVLR aims at finding a low dimensional subspace by subspace learning,and preserves the local geometric structure of feature by combining the feature graph in the algorithm.In addition,GVLR learns a virtual label matrix from the viewpoint of spectral analysis to combine the discriminant learning with unsupervised feature selection.By using the linear regression function,GVLR can bridge the relationship between the feature subspace and virtual label space,and can enhance the discriminant ability of features by fully using the label information.Meanwhile,the-norm is added to improve the robustness of method and sparsity of features.At last,to solve the above-mentioned two methods,an efficient alternative optimization algorithm is utilized.And extensive experiments on several public datasets are conducted with selected features.From the experimental results,the conclusion is that the methods can gain better performance than some classical and state-of-the-art unsupervised feature selection methods.