Abstract—In the last decade, crowdsourcing (CS) has emerged as a very promising approach for obtaining services, feedback or data from a large number of people connected through the Internet, in a short time and at a reasonable cost. CS has been used in a large range of contexts, thus proving its versatility. However, the quality of the services or data provided by the workers (the ”crowd”) is not guaranteed, and therefore must be verified. This verification usually results in additional time and cost. We propose a novel approach of quality control in crowdsourcing that reduces, and in some cases eliminates, this overhead. Our approach uses a learning technique to characterize and cluster tasks, and selects, within the available crowd, the most reliable group of workers for a given type of tasks.