Crowd-Sensing with Polarized Sources

The paper presents a new model for crowd-sensing applications, where humans are used as the sensing sources to report information regarding the physical world. In contrast to previous work on the topic, we consider a model where the sources in question are polarized. Such might be the case, for example, in political disputes and in situations involving different communities with largely dissimilar beliefs that color their interpretation and reporting of physical world events. Reconstructing accurate ground truth is more complicated when sources are polarized. The paper describes an algorithm that significantly improves the quality of reconstruction results in the presence of polarized sources. For evaluation, we recorded human observations from Twitter for four months during a recent Egyptian uprising against the former president. We then used our algorithm to reconstruct a version of events and compared it to other versions produced by state of the art algorithms. Our analysis of the data set shows the presence of two clearly defined camps in the social network that tend of propagate largely disjoint sets of claims (which is indicative of polarization), as well as third population whose claims overlap subsets of the former two. Experiments show that, in the presence of polarization, our reconstruction tends to align more closely with ground truth in the physical world than the existing algorithms.