Feature Selection Based On Bayesian Decision Theory

The minimization of financial risk has always been the focus of investors and researchers.In financial risk management,the two-classification problem is one of the most important problems.Feature selection plays an important role in classification.In this paper,a feature selection algorithm based on Bayes decision rule is proposed.It is expected that the main features affecting the classification results can be screened out and the classification accuracy of the classifier can be greatly improved:(1)Assuming that all features of the sample set are independent,a step-by-step backward algorithm based on Bayes decision rule is proposed.(2)In real life,samples have same label,but some of their attribution are not similar,and these unsimilar attribution will ultimately affect the classification results.Therefore,this paper gives a step-by-step forward algorithm based on LMNN-Bayes decision rule.In order to verify the superiority of the proposed feature selection method,we do simulation.By comparing with the traditional stepwise regression,it is concluded that the new feature selection method has better classification effect and higher sensitivity to noise features.Finally,aiming at the two-classification problem of the investable and non-investable bonds in the Chinese bond market,based on the LMNN-Bayes decision feature selection algorithm and the CVaR minimum two-classification model,this paper applies it to the credit rating data of China's bond market,and compares it with the CVaR minimization two classification model.The former has a better classification effect than the latter,and at the same time we can screens out the important features that affect bond rating,which has a certain practical value.