Supporting code review by automatic detection of potentially buggy changes

Abstract

Code reviews constitute an important activity in software quality assurance. Although they are essentially based on human expertise and scrupulosity, they can also be supported by automated tools. In this paper we present such a solution integrated with code review tools. It is based on a SVM classifier that indicates potentially buggy changes. We train such a classifier on the history of a project. In order to construct a training set, we assume that a change/commit is buggy if its modifications has been later altered by a bug-fix commit. We evaluated our approach on 77 selected projects taken from GitHub and achieved promising results. We also assessed the quality of the resulting classifier depending on the size of a project and the fraction of the history of a project that have been used to build the training set.