Predicting software maintenance effort through evolutionary-based decision trees

Abstract

Software effort prediction has been a challenge for researchers throughout the years. Several approaches for producing predictive models from collected data have been proposed, although none has become standard given the specificities of different software projects. The most commonly employed strategy for estimating software effort, the multivariate linear regression technique has numerous shortcomings though, which motivated the exploration of many machine learning techniques. Among the researched strategies, decision trees and evolutionary algorithms have been increasingly employed for software effort prediction, though independently. In this paper, we propose employing an evolutionary algorithm to generate a decision tree tailored to a software effort data set provided by a large worldwide IT company. Our findings show that evolutionarily-induced decision trees statistically outperform greedily-induced ones, as well as traditional logistic regression. Moreover, an evolutionary algorithm with a bias towards comprehensibility can generate trees which are easier to be interpreted by the project stakeholders, and that is crucial in order to improve the stakeholder's confidence in the final prediction.