Exploiting Class Bias for Discovery of Topical Experts in Social Media

Discovering contexts of user's expertise can be a challenging task, especially if there is no explicit attribution provided. With more professionals adopting social networks as a mean of communicating with their colleagues and broadcasting updates on the area of their competence, it is crucial to detect such individuals automatically. This would not only allow for better follower recommendation, but would also help to mine valuable insights and emerging signals in different communities. We posit that topical groups have their unique semantic signatures. Hence, we can treat identification of expert's topical attribution as a binary classification task, exploiting the class bias to generate training sample without any manual labor. In thiswork, we present profile-and behavior-based models to explore experts topicality. While the former focuses on the static profile of user activity, the latter takes into account consistency and dynamics of a topic in user feed. We also propose a naive baseline tailored to a domain used in evaluation. All models are assessed on a case study of Twitter investment community.