Abstract Software cross-project defect prediction (CPDP) makes use of cross-project (CP) data to overcome the lack of data necessary to train well-performing software defect prediction (SDP) classifiers in the early stage of new software projects. Since the CP data (known as the source) may be different from the new project’s data (known as the target), this makes it difficult for CPDP classifiers to perform well. In particular, it is a mismatch of data distributions between source and target that creates this difficulty. Transfer learning-based CPDP classifiers are designed to minimize these distribution differences. The first Transfer learning-based CPDP classifiers treated these differences equally, thereby degrading prediction performance. To this end, recent research has the Weighted Balanced Distribution Adaptation (W-BDA) method to leverage the importance of both distribution differences to improve classification performance. Although W-BDA has been shown to improve model performance in CPDP and tackle the class imbalance by balancing the class proportion of each domain, research to date has failed to consider model performance in light of increasing target data. We provide the first investigation studying the effects of increasing the target data when leveraging the importance of both distribution differences. We extend the initial W-BDA method and call this extension the W-BDA $$\mathbf {^{+}}$$ + method. To evaluate the effectiveness of W-BDA $$\mathbf {^{+}}$$ + for improving CPDP performance, we conduct eight experiments on 18 projects from four datasets, where data sampling was performed with different sampling methods. Data sampling was only performed on the baseline methods and not on our proposed W-BDA $$\mathbf {^{+}}$$ + and the original W-BDA because data sampling issues do not exist for these two methods. We evaluate our method using four complementary indicators (i.e., Balanced Accuracy, AUC, F-measure and G-Measure). Our findings reveal an average improvement of 6%, 7.5%, 10% and 12% for these four indicators when W-BDA $$\mathbf {^{+}}$$ + is compared to the original W-BDA and five other baseline methods (for all four of the sampling methods used). Also, as the target to source ratio is increased with different sampling methods, we observe a decrease in performance for the original W-BDA, with our W-BDA $$\mathbf {^{+}}$$ + approach outperforming the original W-BDA in most cases. Our results highlight the importance of having an awareness of the effect of the increasing availability of target data in CPDP scenarios when using a method that can handle the class imbalance problem.

Cross-project defect prediction (CPDP) models are used in new software project prediction tasks to improve defect prediction rates. The development of these CPDP models could be challenging in cases where there is little or no historical data. For this reason, researchers may need to rely on multiple sources and use transfer learning-based CPDP for building defect prediction models. These data are typically taken from similar and related projects, but their distributions can be different from the new software project (target data). Although, transfer learning-based CPDP models are designed to handle these distribution differences, but if not correctly handled by the model, may lead to negative transfer. To this end, recent works have focused on building transfer CPDP models, but little is known about how similar or dissimilar sources should be to avoid negative transfer. This paper provides the first empirical investigation to understand the effect of combining different sources with different levels of similarities in transfer CPDP. We introduce the use of the Population Stability Index (PSI) to interpret whether the distribution of the combined or single-source data is similar to the target data. This was validated using an adversarial approach. Experimental results on three public datasets reveal that when the source and target distribution are very similar, the probability of false alarm is improved by 3% to 7% and the recall indicator is reduced from 1% to 8%. Interestingly, we also found that when dissimilar source data are combined with different source datasets, the overall domain divergence is lowered, and the performance is improved. The results highlight the importance of using the right source to aid the learning process.

Stack Overflow is used to solve programming issues during software development. Research efforts have looked to identify relevant content on this platform. In particular, researchers have proposed various modelling techniques to predict acceptable Stack Overflow answers. Less interest, however, has been dedicated to examining the performance and quality of typically used modelling methods with respect to the model and feature complexity. Such insights could be of practical significance to the many practitioners who develop models for Stack Overflow. This study examines the performance and quality of two modelling methods, of varying degree of complexity, used for predicting Java and JavaScript acceptable answers on Stack Overflow. Our dataset comprised 249,588 posts drawn from years 2014-2016. Outcomes reveal significant differences in models’ performances and quality given the type of features and complexity of models used. Researchers examining model performance and quality and feature complexity may leverage these findings in selecting suitable modelling approaches for Q&A prediction.

Context: Stack Overflow is a popular community question and answer portal used by practitioners to solve problems during software development. Developers can focus their attention on answers that have been accepted or where members have recorded high votes in judging good answers when searching for help. However, the latter mechanism (votes) can be unreliable, and there is currently no way to differentiate between an answer that is likely to be accepted and those that will not be accepted by looking at the answer's characteristics. Objective: In potentially providing a mechanism to identify acceptable answers, this study examines the features that distinguish an accepted answer from an unaccepted answer. Methods: We studied the Stack Overflow dataset by analyzing questions and answers for the two most popular tags (Java and JavaScript). Our dataset comprised 249,588 posts drawn from 2014-2016. We use random forest and neural network models to predict accepted answers, and study the features with the highest predictive power in those two models. Results: Our findings reveal that the length of code in answers, reputation of users, similarity of the text between questions and answers, and the time lag between questions and answers have the highest predictive power for differentiating accepted and unaccepted answers. Conclusion: Tools may leverage these findings in supporting developers and reducing the effort they must dedicate to searching for suitable answers on Stack Overflow.