An Improved Deep Forest Regression

Recently deep forest has been modified and applied to regression modeling, namely Deep Forest Regression (DFR). Its results are satisfactory in small sample datasets. However, the diversity of forests is not fully utilized. Therefore, in this paper, an improved DFR (ImDFR) algorithm is proposed to promote regression modeling. With the structural framework unchanged, random forest, completely random forest, GBDT and XGBoost are used as sub-forests at each layer to increase diversity. We applied the proposed method to the high-dimensional and low-dimensional benchmark datasets. Experimental results demonstrate that ImDFR can achieve better prediction results than other approaches, and the results prove that proposed model is effective.