Potential Dependency Analysis Based Task Recommendation Model for Crowdsourcing

With the explosive growth of crowdsourcing tasks, how to design appropriate task recommendation models that help workers to find appropriate tasks and improve task allocation efficiency has become a crucial issue in crowdsourcing. However, most existing task recommendation methods based on the Probabilistic Matrix Factorization (PMF) assume feature vectors are independent of each other, therefore impairing the accuracy of recommendation. In this paper, a potential dependency analysis based task recommendation model (PDA-TR) for crowdsourcing is proposed. Firstly, a worker-task rating matrix is decomposed into a worker potential feature matrix and a task potential feature matrix; secondly, the worker potential feature matrix is decomposed into a worker potential dependency matrix and a feature potential dependency matrix via additive factorization with Gaussian process; thirdly, by using the momentum gradient descent algorithm, the parameter matrices of the proposed model are updated, then a predictive worker-task rating matrix can be calculated. Moreover, when new workers have no ratings to tasks, the proposed model can predict their ratings to tasks through the mined dependencies among workers, which can overcome the complete cold start problem. Based on real-world datasets, experiment results verify that the proposed model has obvious advantages over other baselines in terms of accuracy.