**Appendix A. Additional Details on the Empirical Performance Models Validation**

The validation of different EPM for the set of the atomic models (that was noted in Table 2) is presented in Table A1. *R*<sup>2</sup> and RMSE metrics are used to compare the predictions of EPM and real measurements of the fitting time. The obtained results confirm that the linear EPM with two terms is most suitable for most of the ML models used in the experiments. However, the fitting time for some models (e.g., random forest) is represented better by the more specific EPM. The one-term EPM provides a lower quality than more complex analogs.

**Table A1.** Approximation errors for the different empirical performance models' structures obtained for the atomic ML models. The best suitable structure is highlighted with bold.


The visualization of the performance models predictions for the different cases is presented in Figure A1. It confirms that the selected EPMs allow estimating the fitting time quite reliably.

**Figure A1.** The empirical performance models for the different atomic models: LDA, QDA, Decision Tree (DT), PCA dimensionality reduction model, Bernoulli Naïve Bayes model, logistic regression. The heatmap represent the prediction of EPM and the black points are real measurements.
