Detecting and analyzing fine-grained user roles in social media?

Abstract

While identifying specific user roles in social media -in particular bots or spammers- has seen significant progress, generic and all-encompassing user role classification remains elusive on the large data sets of today?s social media. Yet, such broad classifications enable a deeper understanding of user interactions and pave the way for longitudinal studies, capturing the evolution of users such as the rise of influencers. Studies of generic roles have been performed predominantly in a small scale, establishing fundamental role definitions, but relying mostly on ad-hoc, data set-dependent rules that need to be carefully hand-tuned. We build on those studies and provide a largely automated, scalable detection of a wide range of roles. Our approach clusters users hierarchically on salient, complementary features such as their actions, their ability to trigger reactions and their network positions. To associate these clusters with roles, we use supervised classifiers: trained on human experts on completely new media, but transferable on related data sets. Furthermore, we employ the combination of samples in order to improve scalability and allow probabilistic assignments of user roles. Our evaluation on Twitter indicates that a) stable and reliable detection of a wide range of roles is possible b) the labeling transfers well as long as the fundamental properties don?t strongly change between data sets and c) the approaches scale well with little need for human intervention.