Analysis of Decrease in Accuracy of Two-tier Trees without Using Feature Selection
Sergey A. Saltykov · 2021 14th International Conference Management of large-scale system development (MLSD) · 2021
It has been shown in previous papers that traditional approaches to feature selection can lead to learning a two-tier decision tree, which significantly loses in accuracy to a enumerate of all non-empty subsets of features containing no more than three features. In this paper, on a real, public dataset, we calculated how often such an accuracy loss can occur. It is shown that there is a statistically significant correlation between the coefficient of determination of a two-tier decision tree, constructed without any feature selection, and the magnitude of such a loss in accuracy. The characteristic values for the loss of accuracy in the construction of a two-tier tree without any feature selection, compared to a full enumeration of all nonempty subsets of features containing no more than three features, for a real dataset in some domain are calculated.