文章基本信息

标题：Automatic Identification of Similar Pull-Requests in GitHub’s Repositories Using Machine Learning
本地全文：下载
作者：Hamzeh Eyal Salman ; Zakarea Alshara ; Abdelhak-Djamel Seriai 等
期刊名称：Information
电子版ISSN：2078-2489
出版年度：2022
卷号：13
期号：2
页码：73
DOI：10.3390/info13020073
语种：English
出版社：MDPI Publishing
摘要：Context: In a social coding platform such as GitHub, a pull-request mechanism is frequently used by contributors to submit their code changes to reviewers of a given repository. In general, these code changes are either to add a new feature or to fix an existing bug. However, this mechanism is distributed and allows different contributors to submit unintentionally similar pull-requests that perform similar development activities. Similar pull-requests may be submitted to review in parallel time by different reviewers. This will cause redundant reviewing time and efforts. Moreover, it will complicate the collaboration process. Objective: Therefore, it is useful to assign similar pull-requests to the same reviewer to be able to decide which pull-request to choose in effective time and effort. In this article, we propose to group similar pull-requests together into clusters so that each cluster is assigned to the same reviewer or the same reviewing team. This proposal allows saving reviewing efforts and time. Method: To do so, we first extract descriptive textual information from pull-requests content to link similar pull-requests together. Then, we employ the extracted information to find similarities among pull-requests. Finally, machine learning algorithms (K-Means clustering and agglomeration hierarchical clustering algorithms) are used to group similar pull-requests together. Results: To validate our proposal, we have applied it to twenty popular repositories from public dataset. The experimental results show that the proposed approach achieved promising results according to the well-known metrics in this subject: precision and recall. Furthermore, it helps to save the reviewer time and effort. Conclusion: According to the obtained results, the K-Means algorithm achieves 94% and 91% average precision and recall values over all considered repositories, respectively, while agglomeration hierarchical clustering performs 93% and 98% average precision and recall values over all considered repositories, respectively. Moreover, the proposed approach saves reviewing time and effort on average between (67% and 91%) by K-Means algorithm and between (67% and 83%) by agglomeration hierarchical clustering algorithm.