Model Averaging Of Partial Linear Multinomial Logit Model And Its Application

With the continuous development of artificial intelligence,the problem of how to accurately predict classification has attracted a lot of attention.Real-life samples may originate from complex and skewed distributions,and there is a possibility of nonlinearity in the relationship between explanatory and response variables,resulting in poor prediction using linear models.For better dealing with the classification problem under the influence of nonlinear factors,this thesis proposes a semi-parametric model averaging method to increase the model’s flexibility,reduce the risk of model misspecification,and thus improve classification prediction accuracy.First,this thesis proposes the use of a partial linear multinomial logit model(PLMLM),in which the nonparametric component is estimated using the local partial likelihood method,and the estimation of the unknown coefficients is obtained using the maximum likelihood method.Secondly,the partial linear multinomial logit model averaging(PLMLMA)is established by introducing the model averaging method.In order to extract the information of nonlinear factors adequately but not repeatedly,a single continuous covariate is selected sequentially as the independent variable of the nonparametric function of the PLMLM,so as to construct a series of candidate models with the same number of continuous covariates.Finally,the weights of the candidate models in the model averaging are obtained by using maximum likelihood estimation.In addition,a large number of simulations are made to evaluate the prediction effect of the PLMLMA.In the process of comparison with other methods,it is found that the proposed model can maintain the highest average hit rate while the average mean square error is relatively small.This thesis also verifies the prediction effect of the proposed method using empirical evidence.The first empirical evidence is about wheat variety classification,which is characterized by continuous covariates and strong variable correlation,and the results show that the proposed method in this thesis has a higher average hit rate in the classification prediction.The second empirical evidence is the prediction of body performance categories,where there are discrete covariates and some continuous covariates follow skewed distributions,and the results show that the average hit rate of the proposed method is higher than that of the other methods in the comparison,given the relatively small sample size.