| Pervasiveness and permissiveness of social media made content verification an integral part of any analysis involving online generated data. The most natural and convenient way to assess content trustworthiness and validity is to examine it from the perspective of a user that authored it. With more studies switching their strategies of data collection from event-oriented to user-focused, it is crucial to outline and address the challenges pertaining to this approach. We propose a user-centric analytics paradigm as one of the solutions to this problem. It is casted as a three-tier framework, consisting of detection of topical experts, extraction and interpolation of their opinions and utilization of the latter in social filtering. The first is concerned with automatic identification of user's topical attribution on Twitter: while the platform is so popular among professionals, it does not support an explicit mechanism for community membership. We present three models exploiting semantic signature of a group and examine their performance on a case study of Twitter investment community. Interpolation of missing opinions is intended to handle mass amounts of periods with no activity peculiar to user streams. We introduce a number of community-based models exploiting user's historical activity, content and opinions of his immediate network, and also verify their feasibility to serve as an initialization scheme in low-rank matrix approximation. We also analyze how predictability changes from user to user and build a model based on his characteristics to assess this value beforehand. Finally, we present a concept of social filtering ---an approach with objective to exploit plethora of available historical data for prediction of social trends. In contrast to collaborative filtering, it does not solicit explicit recommendations from users and operates on raw data. It is designed to automatically select "expert" users---individuals whose content is the most reflective of a target central to the application---and transform their posts into predictive signals. |
