ARRAY: Adaptive triple feature-weighted transfer Naive Bayes for cross-project defect prediction

Abstract

Context:Cross-project defect prediction (CPDP) aims to predict defects of target data by using prediction models trained on the source dataset. However, owing to the huge distribution difference, it is still a challenge to build high-performance CPDP models. Objective:We propose a novel high-performance CPDP method named adaptive triple feature-weighted transfer naive Bayes (ARRAY). Methods:ARRAY is characterized by feature weighted similarity, feature weighted instance weight, and the model adaptive adjustment. Experiments are performed on 34 defect datasets. We compare ARRAY with seven state-of-the-art CPDP methods in terms of area under ROC curve (AUC), F1, and Matthews correlation coefficient (MCC) with statistical testing methods. Results:Experimental results show that: (1) on average, ARRAY separately improves MCC, AUC, and F1 over the baselines by at least 18.4%, 6.5%, and 4.5%; (2) ARRAY significantly performs better than each baseline on most datasets; (3) ARRAY significantly outperforms all baselines with non-negligible effect size according to post-hoc test. Conclusion:It can be concluded that: (1) the proposed feature weighted similarity, feature weighted instance weight, and the model adaptive adjustment are very helpful for improving the performance of CPDP models; (2) ARRAY is a more promising alternative for CPDP with common metrics.