Which Is More Important for Cross-Project Defect Prediction: Instance or Feature?

Abstract

Software defect prediction plays an important role in software testing. We can build the prediction model based on historical data. However, for a new project, we cannot be able to build a good prediction model due to lack of historical data. Therefore, researchers have proposed the cross-project defect prediction (CPDP) methods to share the historical data among different projects. In practice, there may be the problems of instance distribution differences and feature redundancy in cross-project datasets. To investigate which is more important for CPDP, instance or feature, we take instance filter and feature selection as examples to show their efficiency for CPDP. Our experiments are conducted on NASA and PROMISE datasets, and the results indicate that feature selection performs better than instance filter in improving the performance of CPDP. We can conclude that feature could be more important than instance for CPDP.