Unsupervised Feature Selection Based On Matrix Factorization And Adaptive Graph

Since unsupervised feature selection algorithms do not require prior category information,they are widely used in the fields of machine learning and pattern recognition.Most of the existing unsupervised feature selection algorithms focus on various forms of data reconstruction and minimize the reconstruction error by discarding features with low contribution.Recent research on unsupervised feature selection suggests that several important information should be retained by selected features.Discriminant information can be obtained by clustering.The local manifold structure can be captured by the learning of embedded graphs.For example,the sample pairwise similarity graph,KNN graph the local linear embedding etc.Many models choose to calculate the similarity between samples in the original feature space of the data.However,there is often a lot of noise and redundancy in the original feature space of the samples.Therefore,it is not appropriate to calculate the internal structure of the data directly in the original feature space.First,an unsupervised feature selection based on robust matrix factorization and adaptive graph(MFAGFS)is proposed.The model can perform robust matrix decomposition,feature selection,and local structure learning under a unified learning framework.Pseudo labels can be obtained by robust matrix decomposition.Pseudo labels and local structure information are used to guide the feature selection process.Then the local structures are adaptively learned from the results of feature selection.Through the interaction of the two basic tasks of local structure learning and feature selection,MFAGFS can accurately capture the structural information of data and select discriminative features.Furthermore,in order to achieve excellent cluster separation,the form of matrix decomposition is improved.Orthogonal basis clustering is achieved by performing orthogonal matrix decomposition in the transformation space,and using robust e2,1 norm as the loss function in the matrix decomposition process.An unsupervised feature selection based on Orthogonal basis Clustering and adaptive graph(OCAGFS)is proposed.In order to verify the performance of the two models proposed in this paper,unsupervised feature selection based on robust matrix factorization and adaptive graph(MFAGFS)and unsupervised feature selection based on orthogonal basis clustering and adaptive graph(OCAGDFS)performs comparative experiments on publicly available datasets.The experimental results demonstrate that the proposed methods outperform many state of the art unsupervised feature selection methods.