Crowdclustering items into overlapping clusters

Crowdclustering clusters data items in a crowdsourcing manner, which makes discovered item categories more consistent with human perception. However, due to diversity of crowdsourcing workers and fluctuation of the number of tasks assigned to each worker, inferring stable and reliable clusters is challenging. Moreover, an item may be associated with multiple attributes, and such items should be put into different clusters, which makes inferring accurate clusters more complicated. To address the challenges above, in this paper we present a robust and fast crowdclustering scheme for finding overlapping clusters of items. Distinguished from existing works, we extract reliable and stable cluster information from workers' answers by majority voting. We then formulate an optimization problem to find overlapping clusters, and develop a nonnegative matrix factorization based approach to approximate the optimal solution. Experiments show the robustness, accuracy and efficiency of our approach.