A Reinforcement Learning Approach To Requirement Tagging For Software Crowdsourcing Projects

Software crowdsourcing platforms which allow public participation attract developers located in different areas to anticipate in software development projects published on these platforms.However,the constant increase in the number of projects in these platforms has cost developers a huge amount of time searching for projects that really suit them.Fortunately,tagging is one of the most effective methods to solve this problem.If every project is tagged according to its requirement content,developers will be able to find projects which suit them well in a shorter time than before and will help them to understand the key points of projects quickly.Thus,this paper proposes a tag recommendation method named“Tag Rec”.Tag Rec models the probability of choosing a specific tag by analyzing the tagging history and uses other two methods named “KGR” and “PMFR” which are also proposed in this paper to generate two independent lists of tags for the project to be recommended.Then a ranking method named “RLR” is proposed to re-rank the tags in the previous step.Thus the highest ranked tags are returned as final recommendations.The contributions of this paper are:(1)Propose a universal probability model construction method.This model can be used to infer the actual probability of tag choosing and will be updated according to the actual choices of tags.(2)Propose tag recommendation algorithm KGR and PMFR.KGR recommends tags based on keywords stored in knowledge graph and the links between those keywords.PMFR is based on the concept of collaborative filtering and recommends personalized tags according to the probability model constructed in this paper.(3)Propose tag re-rank algorithm named RLR which is based on reinforcement learning.This algorithm is designed to re-rank the tag list recommended by KGR and PMFR algorithm according to the tag's short-term and long-term “reward”.Also algorithm PBU is proposed to update model and knowledge graph after recommendation feedback is received.(4)Propose Tag Rec recommendation method which is a combination of KGR,PMFR and RLR algorithm.Finally,this paper conducts a series of experiments on Zhubajie,Jointforce and Movie Lens dataset.The experiments show that Tag Rec recommendation method surpasses other baseline algorithms by 5-10%.KGR algorithm and PMFR algorithm get a 3-5% improvement in precision,10% improvement in coverage though they are about the same with other baseline algorithms in terms of variability metrics.RLR algorithm improves precision and coverage metrics for around 5%.