Flexible Reward Plans for Crowdsourced Tasks

Abstract

We develop flexible reward plans to elicit truthful predictive probability distribution over a set of uncertain events from workers. In general, strictly proper scoring rules for categorical events only reward a worker for an event that actually occurred. However, different incorrect predictions vary in quality, and the principal would like to assign different rewards to them, according to her subjective similarity among events; e.g. a prediction of overcast is closer to sunny than rainy.

We propose concrete methods so that the principal can assign rewards for incorrect predictions according to her similarity between events. We focus on two representative examples of strictly proper scoring rules: spherical and quadratic, where a worker’s expected utility is represented as the inner product of her truthful predictive probability and her declared probability. In this paper, we generalize the inner product by introducing a reward matrix that defines a reward for each prediction-outcome pair. We first show that if the reward matrix is symmetric and positive definite, both the spherical and quadratic proper scoring rules guarantee the maximization of a worker’s expected utility when she truthfully declares her prediction. We next compare our rules with the original spherical/quadratic proper scoring rules in terms of the variance of rewards obtained by workers. Finally, we show our experimental results using Amazon Mechanical Turk.