Effective Solution for Labeling Candidates with a Proper Ration for Efficient Crowdsourcing

One of the core problems of crowdsourcing research is how to reduce the cost, in other words, how to get better results with a limited budget. To save budget, most researchers concentrate on internal steps of crowdsourcing while in this work we focus on the pre-processing stage: how to select the input for crowds to contribute. A straightforward application of this work is to help budget-limited machine learning researchers to get better balanced training data from crowd labeling. Specifically, we formulate the prior information based input manipulating procedure as the Candidate Selection Problem (CSP) and propose an end-squeezing algorithm for it. Our results show that a considerable cost reduction can be achieved by manipulating the input to the crowd with the help of some additional prior information. We verify the effectiveness and efficiency of these algorithms through extensive experiments.