On Factors Selection in CatBoost Models Construction

When solving Machine Learning problems, the construction of models capable of capturing the variability of factors in the training data and then recognizing it when making a prediction is of particular importance. Ensemble models provide such an opportunity. One class of the mentioned models is bousting models, in particular, the CatBoost model. This paper presents an approach to identifying influential inputs of the CatBoost model, provided that factor modifications such as, for example, exponent or logarithm of factors are used. The approach removes the problem of multicollinearity caused by the presence of different representations of the same factor. A numerical example shows the effectiveness of the proposed approach.