Hypergraph Low-rank Feature Selection Multi-target Regression Algorithm

With the development of teach knowledge,lots of researchers proposed variety of algorithms to solve the more and more challenging problems of multi output regression missions.In the regression problem of high-dimension data,multi output regression methods utilize multiple eigen-values of the data to predict multiple output values to achieve better regression performance.In many real applications,the data usually have lots of features,such as the genetic data of biomedical science,the image data in computer vision,the log generated by the web site and the text data of the user's browsing record and et al.In order to improve the utilization efficiency of the high-dimension data and to achieve better classification performance in regression analysis,feature selection for the high-dimension data is necessary to reduce the dimension of high-dimension data.Many researchers proposed low-rand regression model and feature selection and attribute reduction were performed on this basis.For the reason that it is only a simply handle on the data,there is no selection of the sample(removing the noise data and the interference from the group point data to the model),the noise data and the data from the group point will interfere with the model.Therefore,in order to be able to handle high dimensional data multi-output regression problem quickly and steadily,this paper puts forward a multi-output regression algorithm based on the hypergraph and low rank feature selection.Specially,at first,l2,p norm is applied on the basis of basic linear regression for sample selection to reduce the influences of noise and outliers,such that the learned model will have more accurate predictive ability;secondly l2,p norm is applied on the regression coefficient matrix for achieving the global optimal regression coefficient matrix to enhance the robustness of the model;thirdly,the feature selection is performed by the regression coefficient matrix,and then the new feature sample data set is obtained;after that,10-fold cross validation is utilized for model training to learn the final regression model.The regression coefficient matrix which is learned from the model can be used to choose the important features,and the low rank regression can output the close correlation structures between the variables;at last,hypergraph is embedded into the regression model for preserving the local correlation structure of the data,which makes the model much more robust.Experimental results on benchmark dataset show that the proposed method can select the features more effectively than other comparison algorithms,and be applied to the multi-output regression tasks with much better classification performance.