A Reinforcement Learning Solution to Cold-Start Problem in Software Crowdsourcing Recommendations

摘要

Recommendation is one key functionality of software crowdsourcing platforms, which is responsible for recommending developers appropriate software projects, or vice versa. Meanwhile, software crowdsourcing recommendation in practice usually faces a cold-start problem: a platform has not yet gathered sufficient information, and thus its recommendations can be imprecise or unbalanced.To tackle this problem, this paper introduces reinforcement learning into crowdsourcing recommendations, and presents ClusterUCBscRec, a novel project recommending approach to learn user feedbacks actively. ClusterUCBscRec adopts the "explore & exploit" strategy to improve the recommending performance continuously, and therefore goes quickly through the cold-start stage. Besides the project models, developer models built from multiple aspects, including developer profile, preferences and skills are introduced into recommendation. Developers and projects are clustered to speed up training and recommending processes to further improve the performance.We have evaluated ClusterUCBscRec on Jointforce. Experimental results show that the novel approach significantly improves the performance of crowdsourcing recommendations and can solve the cold-start problem effectively, compared with COFIBA and BiUCB.