K-Submodlar Function Based Incentive Mechanisms for Crowd Multi-Labeling

Abstract

Crowd labeling, as a new paradigm of labeling classification problems, has enabled it to create a tremendous amount of high-quality labeling datasets by harnessing extensive ordinary human comprehension at a low cost. However, existing works mainly focus on a single label scene(one instance is only associated with a single label or a category). They do not fit some real applications well where one instance can associate with multiple labels and different categories can have different budget limits. In this paper, we find that the issue can be addressed by introducing $K$ -submodlar function, which has received extensive attention recently. Moreover, we further propose a $K$ -submodlar function based incentive mechanism for crowd multi-labeling scene, satisfying the truthfulness, individual rationality, computational efficiency. Extensive simulations validate the theoretical properties of our mechanism.