Predicting Defectiveness of Software Patches

Abstract

Context: Software code review, as an engineering best practice, refers to the inspection of the code change in order to find possible defects and ensure change quality. Code reviews, however, may not guarantee finding the defects. Thus, there is a risk for a defective code change in a given patch, to pass the review process and be submitted. Goal: In this research, we aim to apply different machine learning algorithms in order to predict the defectiveness of a patch after being reviewed, at the time of its submission. Method: We built three models using three different machine learning algorithms: Logistic Regression, NaÃŕve Bayes, and Bayesian Network model. To build the models, we consider different factors involved in review process in terms of Product, Process and People (3P). Results: Our empirical results show that, Bayesian Networks is able to better predict the defectiveness of the changed code with 76% accuracy. Conclusions: Predicting defectiveness of change code is beneficial in making patch release decisions. The Bayesian Network model outperforms the others since it capturs the relationship among the factors in the review process.