Research On The Cold-start Problem In Software Crowdsourcing Platform

Recently,software crowdsourcing becomes more and more popular among software developers.It can utilize spare resources in the internet and aggregate excellent developers from different location together to co-program.As a special ecommercial platform,there is no doubt that software crowdsourcing platform requires recommender systems.For task assigners,if the platform can recommend proficient developers to them to finish the tasks effectively,these assigners will trust the platform and then assign more tasks.Besides,for developers,if the platform can recommend them right tasks meeting their interests and within their abilities,they will depend the platform,as it can bring a lot of earnings to them.However,building recommender systems for software crowdsourcing platform is not an easy task.Differ to traditional recommender systems,items in software crowdsourcing platform have short life period,which is hard for data accumulation to build an effective recommender system.Therefore,the cold-start problem in software crowdsourcing platform is more severe.Considering this problem,a general resourcing modeling method for resources in software platform is proposed in this paper,and the method can well solve the data heterogeneous problem in different platforms.Then,deep learning techniques are applied to solve the task cold-start problem and genetic programing is applied to solve the developer cold-start problem.The contributions of this paper are listed as following:(1)Although data from different software crowdsourcing platforms is heterogeneous,this paper proposes a general method to model developers and tasks in the platform.In the modeling phase,considering the special features of data from software crowdsourcing,a method called w TF-IDF,which is an extended method of TF-IDF,is proposed.Besides,to better model task popularity,a concept called estimated popularity is proposed.(2)To solve the task cold-start problem,a new multi-modal embedding method,base-ME,is proposed.It is a hybrid recommending algorithm,and can map developer vectors and task vectors to a same space,in which cold tasks can be recommended by inner product.Based on base-ME,a deep-ME method is then proposed,combining deep learning techniques.(3)To solve the user cold-start problem,a topic-based popularity sampling method is proposed.The method applies genetic programming to solve the problem.It can even recommend cold tasks to cold users.Finally,to validate the effectiveness of the proposed algorithms,this paper conducts a serial of experiments.The first experiment is comparing the proposed baseME and deep-ME methods to other state-of-the-art methods.Using three real datasets from software crowdsourcing platforms,our methods get the best result,obtaining 15% lifting on both precision and recall in average.In the second experiment,compared with other methods to solve the user cold-start problem,the topic-based sampling method proposed in this paper has a strong lifting,75% on both precision and recall in average,at the recommendation.