Investigating the use of random forest in software effort estimation

摘要

Over the last two decades, there has been an important increase in studies dealing with the software development effort estimation (SDEE) using machine learning (ML) techniques that aimed to improve the accuracy of the estimates and to understand the process used to generate these estimates. Among these ML techniques, decision tree-based models have received a considerable scholarly attention thanks to their generalization ability and understandability. However, very few studies have investigated the use of random forest (RF) in software effort estimation. In this paper, a RF model is designed and optimized empirically by varying the values of its key parameters. The performance of the RF is compared with that of classical regression tree (RT). The evaluation was performed through the 30% hold-out validation method using three datasets: ISBSG R8, Tukutuku and COCOMO. To identify the most accurate techniques, we used three widely known accuracy measures: Pred(0.25), MMRE and MdMRE. The results show that the optimized random forest outperforms the regression trees model on all evaluation criteria.