Studyon Feature Selection Algorithm Based On Locality Constrained Self-Representation

With the rapid development of information technology,a mass of datawhichcharacterize the less samples and high dimension appear in data mining,machine learning fields.However,it's a great difficulty for the image processing,information retrieval,text classification,pattern recognition problems.Meanwhile,it makes many machine learning methods facing “dimension disaster”.How to obtain useful information from these vast amounts of data has been a challenged and hot research problem.Feature selection methods overcome this problem effectively and achieve the purpose of dimension reduction ultimately.It has been received widely attention in machine learning,pattern recognition,etc.Feature selection chooses the optimal feature subset based on a certain evaluate standard from the original feature set.The purpose of it is that choosing the most representative features from the original data so that data classification or clustering will be effectively improved.Feature selection plays a very important role for the analysis of the data such as text,images and video.Inspired by the theory of feature self-representation,the regularization algorithm(Regularized Self-Representation,RSR)has been an efficient unsupervised feature selection algorithm.However,RSR only takes the self-representation ability of features into account,and neglects the locality structure preserving ability of features,which may deteriorate its performance tremendously.To overcome its limitation,we introduce a local scatter matrix which encodes the locality geometric structure of high-dimensional data to preservethe locality information of the input data.Then,we propose a novel unsupervised feature selection algorithm named Locality Constrained Regularized Self-Representation(LCRSR).Moreover,a simple yet efficient iterative update algorithm is developed to solve LCRSR.To test he effectiveness of the proposed feature selection method,extensive experiments are performed on five publicly available databases(such as JAFFE,ORL,AR,COIL20 and SRBCT).Compared with the state-of-art algorithms,our algorithm obtains a better result.