Combining Multiple Learners Induced on Multiple Datasets for Software Effort Prediction
Ekrem Kocagüneli, Ayşe Bener · 2009
Background: First approaches in software effort prediction depended on regression based models, whereas later models investigated more sophisticated methods like machine learning algorithms. Discussion Points: Single methods or models can discover only a certain part of the high dimensional space of software effort data and a common practice to increase accuracy values is to combine multiple learners. However, merely comparing models over a single dataset on the basis of precision values is not a healthy practice, since each dataset may favor a certain method. Therefore, a solid statistical test is required for comparison. Method: In this study, we adapt a previous study conducted by Khosgoftaar et. al. [1] in the field of software quality analysis to the field of software effort estimation and evaluate our results on the basis of statistical significance tests. Conclusions: Khosgoftaar et. al.’s work[1] was the first of its kind in software quality and we adapted their novel work to software effort prediction. We exploited 14 methods over 3 different software effort prediction datasets under 4 different scenarios and observed similar results to Khosgoftaar et. al., that is multiple learners induced on single dataset do not produce significantly better results.