A Transfer Weighted Homogeneous Ensemble Method for Cross-Project Software Fault Prediction

The selection of learners plays a vital role in constructing a robust and trustworthy software fault prediction (SFP) model. Every learner is distinct in terms of its predictive power. Therefore, choosing the ideal learner becomes difficult, especially for cross-project software fault prediction (CPSFP), which relies on external data (aka reference data) for model building. Although ensemble methods have shown promising results, the recent state-of-the-art is still under-explored. A majority of the proposed ensemble techniques merge the outcome of several diverse base classifiers. Unlike them, to further improve the predictive power, we propose a Transfer Weighted Homogeneous Ensemble Learner (TWHEL), which first assigns the transfer weights to reference data samples and then combines the output of various sub-learners trained by applying the same machine learning algorithm on different folds of the weighted reference data generated using cross-validation. To examine the potency of TWHEL, it is compared with two radical methods: Validation and Voting and ASCI on 21 datasets. We achieve an average improvement of 15.08% to 47.51% and 59.60% to 170.8% w.r.t. MCC and F-measure, respectively, over the compared methods. Further, the results of statistical testing corroborate the findings. We thus conclude that practitioners can create reliable, superior-quality software at a reduced cost using the presented approach.