Nowadays, the wisdom of crowds aids in data labeling via crowd sensors. One of the successful tools worth mentioning is Amazon Mechanical Turk. Uncertainty in crowd labels, however, deteriorates the result of the learning algorithm. In some applications, such as weather and stock forecasting and object tracking, the temporal dependency of data (dynamics) is effective in decreasing label uncertainty. In order to benefit from the existing knowledge in label dynamics, the current study first employs a traditional state-space model and it shows that these models have serious drawbacks, for instance, sensors with a low coverage rate and the existence of a random labeler are the main challenges posed in this process. Then, an appropriate dynamic model for crowd sensors is presented and the Bayesian filter is applied so that true label inference and system parameter learning are performed jointly. The present work will show that the proposed method is robust enough to meet these challenges and performs better in comparison to the existing methods. The results of experiments on synthetic and real data confirm this issue

Keywords