Spotting Tip of the Iceberg in Social Media: Detecting Target Clusters from Indications of Individual Items

Social media users' online activities can be grouped into clusters or events by similar content or common objectives, such as posting using same hashtag, or sharing same post. The targets we want to detect, are a subset of clusters that are relatively rare but can be very influential. Rumors and Persuasion campaigns, are two types of target clusters. In this dissertation, we propose to detect target clusters from indications of individual items. In our framework, we identify and utilize a small proportion of activities that exist only in target clusters. We call these indicating activities signals. By conducting content analysis, we can find signals for different types of target clusters: (1) We begin our work by analyzing asking questions behaviors on Twitter. We find out keywords extracted from those questions can help identify bursting events and emerging information needs. (2) In our work of detecting trending rumors, we find that when there is a rumor, even though most posts do not raise questions about it, there may be a few that do. We adopt this enquiry activity such as questioning the truth of a statement as signal to detect rumors. (3) We propose an algorithm that automatically learns the signals from data. We apply the learned signals to the target-cluster detection framework to detect persuasion campaigns in social media.