Affiliation:
1. Federal Research Center “Computer Science and Control,” Russian Academy of Sciences, 119333, Moscow, Russia
Abstract
A new two-level ensemble regression method, as well as its modifications and application in applied problems, are considered. The key feature of the method is its focus on constructing an ensemble of predictors that approximate the target variable well and, at the same time, consist of algorithms that, if possible, differ from each other in terms of the calculated predictions. The ensemble with the indicated properties at the first stage is constructed through the optimization of a special functional, whose choice is theoretically substantiated in this study. At the second stage, a collective solution is calculated based on the forecasts formed by this ensemble. In addition, some heuristic modifications are described that have a positive effect on the quality of the forecast in applied problems. The effectiveness of the method is confirmed by the results obtained for specific applied problems.
Publisher
The Russian Academy of Sciences
Reference15 articles.
1. Положение о ЦКП “Информатика” // [Электронный ресурс]. Режим доступа http://www.frccsc.ru/ckp (дата обращения 14.02.2023).
2. Zhou Z.H. Ensemble Methods: Foundations and Algorithms. N.Y.: Chapman and Hall/CRC, 2012.
3. Hastie T., Tibshirani R., Friedman J. The Elements of Statistical Learning Data Mining, Inference, and Prediction. Springer Series in Statistics. N.Y.: Springer, 2009.
4. Breiman L. Random forests // Machine Learning. 2001. V. 45. № 1. P. 5–32.
5. Schapire R.E., Freund Y. Foundations and Algorithms. Cambridge, Massachusetts, London: MIT Press, 2012.