On Supervised Deep Gaussian Mixture Models

The paper presents a way to solve supervised learning problems (classification and regression) using deep Gaussian mixture models (DGMMs). In particular, a composition of the classical version of DGMM with supervised learning methods is used. More than 20 datasets with various parameters from the UCI Machine Learning repository were used for testing. It has been demonstrated that the greatest increase in classification accuracy (by 15.01%) is achieved with the combination of DGMM and extreme gradient boosting. DGMM regression (DGMMR) outperformed both the linear regression and the Gaussian mixture regression in terms of RMSE metric by 8.32% and 27.22%, respectively. The ensemble of DGMM classification and extreme gradient boosting also showed the best results in the semi-supervised class, but it is 9.44% inferior to DGMMR in the RMSE metric on test datasets.