Predicting User Roles in Social Networks Using Transfer Learning with Feature Transformation

How can we recognise social roles of people, given a completely unlabelled social network? We may train a role classification algorithm on another dataset, but then that dataset may have largely different values of its features, for instance, the degrees in the other network may be distributed in a completely different way than in the first network. Thus, a way to transfer the features of different networks to each other or to a common feature space is needed. This type of setting is called transfer learning. In this paper, we present a transfer learning approach to network role classification based on feature transformations from each network's local feature distribution to a global feature space. We implement our approach and show experiments on real-world networks of discussions on Wikipedia as well as online forums. We also show a concrete application of our approach to an enterprise use case, where we predict the user roles in ARIS Community, the online platform for customers of Software AG, the second-largest German software vendor. Evaluation results show that our approach is suitable for transferring knowledge of user roles across networks.