Code Reviewer Intelligent Prediction in Open Source Industrial Software Project

Currently, open-source software is gradually being integrated into industrial software, while industry protocols in industrial software are also gradually transferred to open-source community development. Industrial protocol standardization organizations are confronted with fragmented and numerous code PR (Pull Request) and informal proposals, and different workflows will lead to increased operating costs. The open-source community maintenance team needs software that is more intelligent to guide the identification and classification of these issues. To solve the above problems, this paper proposes a PR review prediction model based on multi-dimensional features. We extract 43 features of PR and divide them into five dimensions: contributor, reviewer, software project, PR, and social network of developers. The model integrates the above five-dimensional features, and a prediction model is built based on a Random Forest Classifier to predict the review results of PR. On the other hand, to improve the quality of rejected PRs, we focus on problems raised in the review process and review comments of similar PRs. We propose a PR revision recommendation model based on the PR review knowledge graph. Entity information and relationships between entities are extracted from text and code information of PRs, historical review comments, and related issues. PR revisions will be recommended to code contributors by graph-based similarity calculation. The experimental results illustrate that the above two models are effective and robust in PR review result prediction and PR revision recommendation.

Automated summary

This paper proposes a PR review prediction model based on multi-dimensional features and a PR revision recommendation model based on the PR review knowledge graph, in order to solve the code review problems and improve the quality in the open-source community. By extracting and classifying the 43 features of PR, a prediction model based on Random Forest Classifier is built to predict the review results. Meanwhile, using graph-based similarity calculation, PR revisions are recommended based on historical review comments and related issues. The experimental results demonstrate the effectiveness and robustness of these two models in PR review and revision.