Study On Unsupervised Feature Selection Algorithm Based On Regularized Matrix Factorization

With the rapid development of science and technology,vast amounts of data which usually characterize the diversity,high dimension and redundancy appear in machine learning,data mining fields.It makes many machine learning methods face serious “Curse of Dimensionality”.As a result,researching the algorithms to process a mass of these initial data becomes a research focus.Among them,the simple and efficient of feature selection method has received the widespread attention.It has important research significance and wide practical application in pattern recognition and scientific research fields.Feature selection is intended to select a subset of features which is strongly correlated with the category and weakly correlated between the features.Thereby the performances of the algorithm model will be effectively improved for data classification or clustering.Recently,inspired by matrix factorization,Matrix Factorization Feature Selection?MFFS?has been efficiently applied to image recognition.MFFS treated feature selection as a matrix factorization problem and successfully constructed a bridge between the matrix factorization and feature selection.However,the orthogonal constraint is too strict to be satisfied in practice and the correlations among features in also neglected in it,which may restrict the overall performance of the method.To address the limitations of the above algorithm,we introduce the inner product regularization to considerate the correlation among features.We first propose a novel unsupervised feature selection algorithm named Regularized Matrix Factorization Feature Selection?RMFFS?.Selected feature subset by our method can not only approximately represent all original high-data,but also low redundancy.Meanwhile,a simple yet efficient iterative optimization algorithm is proposed to solve RMFFS.In order to verify the effectiveness of the proposed method,we compare our approach with other state-of-the-art unsupervised algorithms on six standard databases?such as AR10 P,Yale,ORL,Jaffe,PIE10 P and TOX₁71?.Extensive experimental results show that the performance of our algorithm is superior to others.