Background: Cross project defect prediction (CPDP) recently gained considerable attention, yet there are no systematic efforts to analyse existing empirical evidence. Objective: To synthesise literature to understand the state-of-the-art in CPDP with respect to metrics, models, data approaches, datasets and associated performances. Further, we aim to assess the performance of CPDP versus within project DP models. Method: We conducted a systematic literature review. Results from primary studies are synthesised (thematic, meta-analysis) to answer research questions. Results: We identified 30 primary studies passing quality assessment. Performance measures, except precision, vary with the choice of metrics. Recall, precision, f-measure, and AUC are the most common measures. Models based on Nearest-Neighbour and Decision Tree tend to perform well in CPDP, whereas the popular naïve Bayes yields average performance. Performance of ensembles varies greatly across f-measure and AUC. Data approaches address CPDP challenges using row/column processing, which improve CPDP in terms of recall at the cost of precision. This is observed in multiple occasions including the meta-analysis of CPDP versus WPDP. NASA and Jureczko datasets seem to favour CPDP over WPDP more frequently. Conclusion: CPDP is still a challenge and requires more research before trustworthy applications can take place. We provide guidelines for further research.

Context: Search based approaches are gaining attention in cross project defect prediction (CPDP). The complexity of such approaches and existence of various design decisions are important issues to consider. Objective: We aim at investigating factors that can affect the performance of search based selection (SBS) approaches. We study a genetic instance selection approach (GIS) and present an evaluation of design options for search based CPDP. Method: Using an exploratory approach, data from different options of models are gathered and analyzed through ANOVA tests and effect sizes. Results: Both feature sets and validation dataset selection options show small or insignificant impacts on F-measure and precision, unlike the more affected false positive and true negative rates. Size of training data does not seem to be related to significant changes in F-measure and precision and high variability in performance are discouraging evidence for using larger datasets. Fitness function is one of the major factors that impact performance with much larger effect than the choice of validation dataset. Finally, while showing slight impacts, data label changes do not seem to be the top contributor to performance. Conclusions: We conclude that exploratory approaches can be effective for making design decisions in constructing search based CPDP models. Effect of individual tuned learners and their interaction with other affecting parameters and more in depth study of quality affecting factors guided by label changes are directions to investigate.