Hyper-parameter tuning of classification and regression trees for software effort estimation

Abstract

Classification and regression trees (CART) have been reported to be competitive machine learning algorithms for software effort estimation. In this work, we analyze the impact of hyper-parameter tuning on the accuracy and stability of CART using the grid search, random search, and DODGE approaches. We compared the results of CART with support vector regression (SVR) and ridge regression (RR) models. Results show that tuning improves the performance of CART models up to a maximum of 0.153 standardized accuracy and reduce its stability radio to a minimum of 0.819. Also, CART proved to be as competitive as SVR and outperformed RR.