Software Effort Estimation Using an Optimal Trees Ensemble: An Empirical Comparative Study

Since information systems have become the heartbeat of many organizations, the investment in software is growing rapidly and consuming then a significant portion of the company budget. In this context, both the software engineering practitioners and researchers are more interested than ever about accurately estimating the effort and the quality of software product under development. Accurate estimates are desirable but no technique has demonstrated to be successful at effectively and reliably estimating software development effort. In this paper, we propose the use of an optimal trees ensemble (OTE) to predict the software development effort. The ensemble employed is built by combining only the top ranked trees, one by one, from a set of random forests. Each included tree must decrease the unexplained variance of the ensemble for software development effort estimation (SDEE). The effectiveness of the OTE model is compared with other techniques such as regression trees, random forest, RBF neural networks, support vector regression and multiple linear regression in terms of the mean magnitude relative error (MMRE), MdMRE and Pred(l) obtained on five well known datasets namely: ISBSG R8, COCOMO, Tukutuku, Desharnais and Albrecht. According to the results obtained from the experiments, it is shown that the proposed ensemble of optimal trees outperformed almost all the other techniques. Also, OTE model outperformed statistically the other techniques at least in one dataset. © 2020, Springer Nature Switzerland AG.