Research On Gray Sheep Problem In Project Recommendation System Of Software Crowdsourcing

In recent years,due to the rapid development of software crowdsourcing,a large number of users and projects are published on the software crowdsourcing platforms.To help users quickly find the most suitable projects,academia and industry study the software crowdsourcing platform recommender system,one of the most commonly used recommendation algorithm is collaborative filtering(CF).However,CF is facing greysheep user problem.The "Gray-Sheep" means that users whose preference is different from most persons.Therefore,CF cannot recommend suitable projects to grey-sheep users as it tends to recommend popular projects.In order to solve the problem of grey-sheep users,this paper analyzes the distribution of them and their impact on the recommendation system,then designs algorithms to identify grey-sheep users,and proposes two recommendation approaches for grey-sheep users.The first approach is based on identification of grey-sheep users and white-sheep users,respectively,using two hybrid recommendation models,and it is applicable for platforms that the account of users often changes,or demand high interpretability.The second approach uses a Stacked AutoEncoder,to recommend projects by deep collaborative filtering recommendation,and it is applicable for platforms that the account of users is stable,or require high interpretability results.The main contributions and innovations in this paper include:(1)Construct user's portrait of software crowdsourcing platform,calculate user's basic feature vector and interest feature vector,and analyze multi-type user's behavior to construct implicit rating matrix.A grey-sheep user identification algorithm that can be used in many scenarios is proposed,including neighborhood-based,density-based,angle-based,linear-based model and nonlinear-based model.(2)A project recommendation approach based on hybrid recommendation is proposed,which combines weighted CF,content-based recommendation and graphbased recommendation.And it then transforms the hybrid recommendation list integration into supervised training problem.(3)A project recommendation approach based on deep CF recommendation is proposed,which embeds user portrait,project portrait and implicit rating matrix into implicit factor space.Combining matrix factorization and Stacked AutoEncoder makes it possible to directly calculate the forecast rating between users and projects.This paper conducts a series of experiments on the data sets of multiple software crowdsourcing platforms.The experimental results show that gray-sheep users account for 15%-25% of the total number of users in the crowdsourcing platform,which makes the CF recommendation suboptimal results.In this paper,the project recommendation algorithms based on hybrid recommendation and the algorithm based on deep CF are proposed.By experimental results on three different datasets,the average percentage increases in R@K for grey-sheep users is 10.74% and 12.63%,and the average percentage increases in R@K for all users is 56.65% and 48.25%.