MODEL-BASED SOFTWARE EFFORT ESTIMATION - A ROBUST COMPARISON OF 14 ALGORITHMS WIDELY USED IN THE DATA SCIENCE COMMUNITY

The emergence of the data science discipline has facilitated the development of novel and advanced machine-learning algorithms for tackling tasks related to data analytics. For example, ensemble learning and deep learning have frequently achieved promising results in many recent data-science competitions, such as those hosted by Kaggle. However, these algorithms have not yet been thoroughly assessed on their performance when applied to software effort estimation. In this study, an assessment framework known as a stable-ranking-indication method is adopted to compare 14 machine-learning algorithms widely adopted in the data science communities. The comparisons were carried out over 13 industrial datasets, subject to six robust and independent performance metrics, and supported by the Brunner statistical test method. The results of this study proved to be stable because similar machine-learning algorithms achieved similar performance results; particularly, random forest and bagging performed the best among the compared algorithms. The results further offered evidence that demonstrated how to build an effective stacked ensemble. In other words, the optimal approach to maximizing the overall expected performance of the stacked ensemble can be derived through a balanced trade-off between maximizing the expected accuracy by selecting only the solo algorithms that are most likely to perform outstandingly on the dataset, and maximizing the level of diversity of the algorithms. Precisely, the stack combining bagging, random forests, analogy-based estimation, adaBoost, the gradient boosting machine, and ordinary least squares regression was shown to be the optimal stack for the software effort estimation datasets.