arXiv2015
Given the ever increasing amount of publicly available social media data, there is growing interest in using online data to study and quantify phenomena in the offline "real" world. As social media data can be obtained in near real-time and at low cost, it is often used for "now-casting" indices such as levels of flu activity or unemployment. The term "social sensing" is often used in this context to describe the idea that users act as "sensors", publicly reporting their health status or job losses. Sensor activity during a time period is then typically aggregated in a "one tweet, one vote" fashion by simply counting. At the same time, researchers readily admit that social media users are not a perfect representation of the actual population. Additionally, users differ in the amount of details of their personal lives that they reveal. Intuitively, it should be possible to improve now-casting by assigning different weights to different user groups. In this paper, we ask "How does social sensing actually work?" or, more precisely, "Whom should we sense--and whom not--for optimal results?". We investigate how different sampling strategies affect the performance of now-casting of two common offline indices: flu activity and unemployment rate. We show that now-casting can be improved by 1) applying user filtering techniques and 2) selecting users with complete profiles. We also find that, using the right type of user groups, now-casting performance does not degrade, even when drastically reducing the size of the dataset. More fundamentally, we describe which type of users contribute most to the accuracy by asking if "babblers are better". We conclude the paper by providing guidance on how to select better user groups for more accurate now-casting.