Adopting Learning-to-rank Algorithm for Reviewer Recommendation

Software code review is a software quality assurance activity in which one or several developers check the quality of a source code change. The success of a code review depends on finding appro­priate reviewers to inspect the code change, e.g., a pull request (PR). Otherwise, it can result in inefficient or low-quality code reviews. To match the expertise of a reviewer with a PR, existing approaches model the expertise of reviewers using different features (e.g., the file path similarity, the textual similarity, and the social connection features). However, the weights of different features are usually handcrafted and customized for each project. It is often time-consuming as the weights used in one project cannot be propa­gated to other projects. In this paper, we propose a learning-to-rank (LtR} approach that can automatically learn the best weights of features to recommend reviewers. We empirically study 80 GitHub projects to test the performance of our LtR approach and compare its performance with two baselines. The experiment results show that: (1) applying the maximum aggregation scheme to compute features improves the performance of our LtR approach; (2) the LtR approach outperforms the two baseline models by 28%, 10%, and 12% with respect to top-1 and top-3 accuracy, and Mean Reciprocal Rank on average; (3) the semantic similarity feature can be used to recommend appropriate reviewers; (4) and the social connection between the contributors and the reviewers is the most important feature to recommend appropriate reviewers to review PRs.