Meta Learning and Software Effort Estimation

Abstract

© 2020 IEEE. Studies in software effort estimation typically seek to discover the best estimator which will perform best in any circumstances. However, the present study opposes that such a single best estimator may not exist, and rather suggests to seek the solution to determine the most suitable estimator for a given dataset. This study proposes MLS (short for a meta-learning framework for software effort estimation) to explore the possibility to apply meta learning to the software effort estimation problem. The approach hypothesizes that the same estimation algorithm will offer similar accuracy for similar datasets, where the similarity between datasets is determined by quantifying their meta characteristics, such as the mean absolute correlation between all the numeric variables. Based on the performance evaluation undertaken on ten industrial datasets, together with six robust performance measures, and Brunner's robust statistical test method, MLS outperforms 48 other effort estimators. It can be preliminary concluded that the meta-learning approach developed in the study is able to provide higher accuracy than to use only one single estimator on all the datasets. More studies on meta learning would be greatly useful to establish a greater degree of accuracy on the estimation activities in empirical software engineering.