Research On Graph Embedding Feature Selection Model Based On Sparse Constraints

Feature selection model based on sparse model and graph embedding attracts much attention in recent years.It has been widely applied in many fields such as machine learning and pattern recognition.The study shows that sparse constraints and graph embedding can effectively improve the robustness and learning efficiency of the feature selection model,help to select efficient features,and boost the accuracy of image classification.In order to improve the robustness and learning efficiency of the feature selection model,several new feature selection models based on sparse constraints and graph embedding are proposed,with unsupervised learning and semi-supervised learning as the theoretical background.Specifically,the main research contents of this paper are as follows:(1)The traditional feature selection model based on graph is mostly based on two independent steps: defining the graph structure in advance and then learning the projection matrix.This makes the feature selection effect highly dependent on the quality of the initial definition of the graph.Because the initial definition of the graph may not be optimal,the algorithm is difficult to achieve the optimal.To solve this problem,a new unsupervised feature selection model based on local and similarity preserving is proposed.This model integrates graph learning and sparse reconstruction into the same framework for learning,allowing the coding process to be sparse,adaptive neighbor and non-negative,ensuring that the optimal graph structure is learned.By introducing the sparse 2,1l-norm of row consistency,the selected features can not only maintain the locality of data and the similarity between reconstruction coefficient,but also have strong discrimination and stability,thus boosting the robustness and learning efficiency of the model.(2)Aiming at the shortcoming of traditional feature selection model based on graph,that is,the two-step strategy of first defining the graph and then learning the projection matrix cannot make the algorithm optimal,a new unsupervised feature selection model based on local preserving projection is proposed.Specifically,the model integrates graph learning and projection learning into the same frame,learns the similarity matrix adaptively through sparse representation and local coordinate coding.At the same time,the local structure of data is well preserved by similar graphs embedded,and then the projection matrix learned is more suitable for feature selection.(3)Aiming at the shortcomings of traditional semi-supervised feature selection method based on graph,that is,the predefined graph cannot accurately guide the transfer of label information,and it is difficult to achieve the overall optimization,a semi-supervised feature selection model based on non-negative sparse graph is proposed.This model integrates label prediction and projection learning into linear regression,and then unifies linear regression and graph structure learning in the same frame,so that linear regression and graph learning can be carried out at the same time to ensure the overall optimization.The graph learned can accurately transmit label information,which also enables linear regression to learn a more discriminant projection,so as to better fit sample labels and accurately classify new samples.In summary,three feature selection models based on sparse constraints and graph embedding are proposed in this paper.Through theoretical analysis and experiments on image classification tasks,it is proved that the proposed model can effectively select efficient feature subset,improve the effect of subsequent data analysis,and has high robustness and learning efficiency.