The Support Vector Machine Classification Model Based On Weighted Punishment Method And Its Application

The rapid development of information technology has spawned a large number of machine learning algorithms.Support vector machine(SVM)is widely used as a theoretical machine learning method.However,this method also has the defect of traditional machine learning methods,that is,as a black box model,it cannot screen out important variables in the model.In order to improve the interpretability of the model,scholars combine the support vector machine model with the penalty function with the function of variable selection.On the basis of previous studies,this paper will add quadratic penalty to the elastic network model and combine the weighted elastic network and support vector machine model to construct the weighted elastic network support vector machine model.According to the different weight construction methods,it can be divided into the following types of models :SVM-RF-ENet,SVM-Lasso-ENet,SVM-Ridge-ENet,SVM-Ad-ENet,SVM-Graph-ENe t,SVM-RF Graph-ENet.In this paper,semi-smooth Newton coordinate descent(SNCD)is used to solve the optimization problem of the target model.In addition,AUC and Recall values were used as evaluation criteria,and the optimal parameters were selected by 10-fold cross validation.In the three different simulation data set,we compare the prediction accuracy and variable selection effect between the weighted elastic network support vector machine model and the elastic network support vector machine model.The results show that the weighted method has significant advantages in variable selection,and some of the weighted methods have improved prediction accuracy,especially in the high-dimensional correlation data.In order to verify the effect of the model in the empirical data,this paper applies the weighted elastic network support vector machine model to the financial distress prediction of listed companies.The sample data are the a-share companies in financial distress from 2018 to 2021 as the positive sample,and the remaining A-share companies as the negative sample.We construct balanced data set by random sampling of negative samples and use bagging method to model.It is concluded that the weighted method shows obvious advantages in empirical data.Most of the models have improved significantly in variable selection and prediction accuracy compared with unweighted methods,among which SVM-RF-Graph-ENet has the best performance.The model simplifies the model as much as possible on the basis of achieving high prediction accuracy,and makes the model close to the reality.The results of comprehensive variable selection show that the ratio structure class,the index class per share and the index of profitability which can reflect a company’s asset structure,growth ability and income quality are of significant significance to the prediction of corporate financial distress.