Machine learning embedded EM algorithms for semiparametric mixture regression models

In this article, we propose two machine learning embedded algorithms for a class of semiparametric mixture models, where the mixing proportions and mean functions are unknown but smooth functions of covariates. Embedding machine learning techniques into a modified EM algorithm, the hybrid estimation technique applies the neural network to estimate the nonparametric parts of the model while keeping the structure of the mixture regression model. Compared to the kernel-based techniques, the new method greatly improves the estimation of the nonparametric functions, when the dimension of the covariates is moderately high. Simulation and real data analysis show the superiority of the new method.