Abstract
Detection of human behavior in On-line Social Networks (OSNs) has become a very important challenge for a wide range of appli- cations, such as security, marketing, parent controls and so on, opening a wide range of novel research areas, which have not been fully addressed yet. In this paper, we present a two-stage method for finding unexplained (and potentially anomalous) behaviors in social networks. First, we use Markov chains to automatically learn from the social network graph a number of models of human behaviors (normal behaviors); the second stage applies an activity detection framework based on the concept of possible words to detect all unexplained activities with respect to the well-known behaviors. Some preliminary experiments using Facebook data show the approach efficiency and effectiveness.