Research On A Reviewer Recommender From An Industrial Perspective

As an important component of software quality assurance,code review aims to secure code quality,which has been increasingly emphasized in modern software development.However,with the developing the large scale software development,how to find the suitable reviewer in a timely manner from the large number of relevant software developers has become an important issue in code review practice.There are two main problems in the current code reviewer recommendation practice research.On the one hand,most studies focused on the application in open source projects,and few of them investigated code review practice in industry.On the other,most existing study utilize recommending rules to select reviewers.Although some research tried to use machine learning algorithms to integrate multi-dimensional features in recommendation practice,some important and commonly used recommendation rules are still difficult to integrate into the machine learning models.Based on a series inquiries in enterprise about their code review process,reviewer selection strategy and relevant data,we proposed a syncretic reviewer recommendation framework,named FPS-RF Recommender.The framework extracts features from the perspectives of reviewer,code and file path,and use Random Forest as classification algorithm.The framework merged the traditional file path similarity based recommender with machine learning model.This framework uses TF-IDF algorithm to transform file path similarity into a feature with finite dimensions.In order to verify the validity of the framework,this thesis selects eight enterprise projects,and uses the traditional file path similarity method and machine learning methods based on decision tree,support vector machine and simple bayesian algorithms for comparison.It was found that the FPS-RF Recommender showed the best performance among all the methods.The highest accuracy of the framework is 87.8%in Top-5 and 78.7% in Top-1.At the same time,in order to improve the recommendation performance of the framework in practice,this thesis proposes a reverse time series validation model,which verifies the influence of the size of training set on reviewer recommendation.The experiment shows that the review records in the recent time span as the training set can improve the performance of the recommendation.At the same time,combining with enterprise research and experimental analysis,this thesis discusses the particularity and some considerations of reviewer recommendation in enterprise scenarios from the perspectives of customized review process,data availability and personnel change,which provides reference value for future enterprise practice.