Research On The Feature Selection Method Based On Multi-sparse Hypergraph Learning Model

As the processing core of human functions,the brain has the most complex physiological structure.Combining functional brain network methods with machine learning has become a very effective research method in the field of neuroimaging.In the research process of brain network classification of magnetic resonance images,due to the rapid development of storage technology and computing power,the scale of brain network and the calculated feature dimensions are very large,and the classification process often causes a lot of damage to the classifier.A lot of pressure,and it is easy to cause over-fitting problems.Therefore,feature selection is very important before classification.In the past research on feature selection,the high-order relationship between feature samples is often ignored,that is,the high-order structure information between multiple samples.Some studies have proposed to use the hypergraph method to model the high-level structural information among the subjects,but this method ignores the influence of the hypergraph generation method on the information representation in the construction process.The hypergraph generated by the K neighboring algorithm lacks global information and group effects,which reduces the effectiveness of feature selection and ultimately leads to a decrease in classification accuracy.In order to solve this problem,this paper proposes to use the Least absolute shrinkage and selection operator(Lasso)to solve the hypergraph generated by sparse linear regression to construct the Laplacian regularization term of the hypergraph.This solves the problem that the global information cannot be effectively represented in the feature selection process,but the grouping effect information is still not well represented.Therefore,it is further proposed to use the Elastic net method and the Group lasso method to generate the hypergraph,and establish the Laplacian regularization term of the hypergraph.This further makes up for the interpretation of the grouping effect information in the feature selection process.The main research work of this paper includes:Propose a variety of methods to solve the sparse linear regression problem to generate the hypergraph,and build the regularization term of the hypergraph on the basis of the novel hypergraph,and then obtain the sparse hypergraph feature learning model.According to the different methods of hypergraph construction,the methods proposed in this paper can be divided into three types.In addition,in order to preliminarily verify the three methods proposed in this paper,this paper selects eight data sets in the UCI database for pre-experiment.The experimental results can preliminarily show the effectiveness of the methods proposed in this paper.This laid the foundation for the subsequent research on mental illness,and after reviewing the follow-up experiment and then looking at the preliminary experiment,it can be found that the method proposed in this article has a certain degree of scalability.Apply the three methods mentioned in the self-collected data set of resting state MRI for depression to realize feature selection and classification model construction.Perform classification on three classifiers,and finally have the best classification effect in the support vector machine of the RBF kernel.Compared with traditional methods,the sparse hypergraph feature learning method proposed in this paper can select more discriminative features and achieve higher classification accuracy.In addition,to verify the influence of the template used on the experiment,five different node definition templates were applied to the self-collected depression data set,and the results all have higher classification accuracy than the other three methods.These results all show that the proposed method deeply mines the potential high-level information among the subjects,which includes global information and group effect information.Relief weight evaluation is carried out on the selected different characteristics,and the newly proposed method is improved compared with the traditional method.Apply the three methods proposed in the self-collected data set of Eysenck personality diffusion tensor to realize feature selection and classification model construction.The experimental results of the three classifiers show that the method proposed in this article has a higher classification accuracy than the traditional method,indicating that the features selected by the method in this article have higher discriminativeness.The hypergraph constructed by different methods to solve the sparse linear regression problem has a greater advantage than the traditional K neighbor algorithm based on Euclidean distance.Under the three newly proposed methods,the sparse hypergraph learning method based on Elastic net and Group lasso method has higher classification accuracy than that based on Lasso in most cases,which shows that the hypergraph generated based on Elastic net contains It contains more potential information than the hypergraph generated based on Lasso,such as grouping effect information.The results also show the effectiveness of the method proposed in this paper in the study of Eysenck personality classification based on diffusion tensor imaging.