Enhancing the performance of gradient boosting trees on regression problems

Abstract Gradient Boosting Trees (GBT) is a powerful machine learning technique that is based on ensemble learning methods that leverage the idea of boosting. GBT combines multiple weak learners sequentially to boost its prediction power proving its outstanding efficiency in many problems, and hence it is now considered one of the top techniques used to solve prediction problems. In this paper, a hybrid approach is proposed that combines GBT with K-means and Bisecting K-means clustering to enhance the predictive power of the approach on regression datasets. The proposed approach is applied on 40 regression datasets from UCI and Kaggle websites and it achieves better efficiency than using only one GBT model. Statistical tests are applied, namely, Friedman and Wilcoxon signed-rank tests showing that the proposed approach achieves significant better results than using only one GBT model.