On the impossibility of convex inference in human computation

摘要: Human computation or crowdsourcing involves joint inference of the ground-truth-answers and worker-abilities by optimizing an objective function, for instance, maximizing data likelihood based on assumed underlying model. A variety methods have been proposed in literature to address this problem. As far as we know, none functions existing is convex. In machine learning applied statistics, a convex function such support vector machines (SVMs) generally preferred, since it can leverage high-performance algorithms rigorous guarantees established extensive optimization. One may thus wonder if there exists meaningful problem human computation. paper, investigate convexity issue We take axiomatic approach formulating set axioms that impose two mild natural assumptions inference. Under these axioms, show unfortunately impossible ensure On other hand, interestingly, absence requirement model "spammers", one construct reasonable guarantee