Feautre stability based machine learning model selection method with usage of Shapley values

Abstract

In this article the usage of Shapley vectors in regression analysis as a method to reduce destabilizing impact of feature multicollinearity and its usage in interpreting machine learning results is considered. The limitations of its application were defined. A Shapley value based method of stable model selection, allowing for stabilization of models’ precision in event of feature and noise distortion, and for increasing precision of classic and innovative ensemble algorithms while shrinking the dataset is proposed. The developed algorithm was tested on both synthetic and publicly available popular machine learning datasets with different amounts of attributes and observation periods. The experiments showed a stable positive effect of decreasing WMAPE and increasing of the effect upon increasing the feature amount of sampling. The suggested algorithm can be used as a tool to increase the efficiency of the ensemble machine learning algorithms, including the high-speed and high-efficiency algorithms.