Role-Based Clustering For Collaborative Recommendations In Crowdsourcing System

Crowdsourcing as a distributed problem-solving and business production model has attracted much attention in recent years. In crowdsourcing systems, task recommendation can help workers to select suitable tasks on crowdsourcing platforms as well as help requesters to receive good outputs. However, as one of the most successful recommendation approaches, current clustering-based models in crowdsourcing are challenged by multi-preference and cold-start problems. This paper proposes a role-based clustering model, which transforms a large-sparse worker-task rating matrix into a set of role-based clusters that are small, independent and rating intensive worker-task rating matrices, leading to better quality and performance in task recommendation. Specifically, we first cluster a worker-task rating matrix into a set of clusters in terms of the role identification and distribution operations. The clusters are further extended to include all their external worker (task) roles. Then, the task recommendation results with respect to a worker are generated by operating over the clusters involving the worker's activities, which captures the worker's preferences in multiple areas. Moreover, the model discovers the structure information from the clustering results and crowdsourcing datasets, by which tasks can be recommended to new workers interactively without their interest profiles. We evaluated our method over the benchmark dataset from NAACL 2010 workshop. The results show the high superiority of our proposed recommendation method over crowdsourcing platforms.