A decision tree is a machine learning algorithm featuring a number of advantages (good interpretability, the ability to work with different types of data, automatic rule generation, etc.). In some tasks, where the data set is limited and attributes are interrelated (e.g., in the text attribution task), linear combinations of attributes can be used instead of single attributes when constructing conditions at the model’s nodes. Taking this into account, the article presents an advancement of the decision tree method, which consists in supplementing k old attributes with new attributes as linear combinations of two original ones: xij = αxi±(1−α)xj , where i, j = 1, . . . , k and the parameter α ∈ [0, 1]. To conduct numerical experiments,the construction of linear combinations of attributes was implemented in the SMALT information system («Statistical Methods for the Analysis of Literary Text») for the case of the text attribution task. The results of classifying pre-1917 texts from the «Vremya» and«Epokha» magazines, and the weekly edition «Grazhdanin» using different types of n-grams showed that this improvement increases the accuracy of the method, but does not significantly reduce the interpretation of the outcome.