Learn Together, Stop Apart: a Novel Approach to Ensemble Pruning

Gradient boosting is the most popular method of constructing ensembles that allow getting state-of-the-art results on many tasks. One of the critical parameters affecting the quality of the learned model is the number of models in the ensemble, or the number of boosting iterations. Unfortunately, the problem of selecting the optimal number of models still remains open and understudied. In this paper, we propose a new look at the hyperparameter selection problem in ensemble models. In contrast to the classical approaches that select the universal size of the ensemble from a hold-out validation subsample, our algorithm uses the hypothesis of heterogeneity of the sample space to adaptively set the required number of steps in one common ensemble for each group of objects individually. Experiments on popular implementations of gradient boosting show that the proposed method does not affect the complexity of learning algorithms and significantly increases quality on most standard benchmarks up to 1.5\%.