A framework for validating the merit of properties that predict the influence of a twitter user

What characterizes an influential user? While there is much research on finding the concrete influential members of a social network, there are less findings about the properties distinguishing between an influential and a non-influential user. A major challenge is the absence of a ground truth, on which supervised learning can be performed. In this study, we propose a complete framework for supervised separation between influential and non-influential users in a social network. The first component of our framework, the InfluenceLearner, extracts a Relation Graph and an Interaction Graph from a social network, computes network properties from them and then uses them for supervised learning. The second component of our framework, the SNAnnotator, serves for the establishment of a ground truth through manual annotation of tweets and users: it contains a crawling mechanism that produces a batch of tweets to be annotated offline, as well as an interactive interface that the annotators can use to acquire additional information about the users and the tweets. On this basis, we have created a ground truth dataset of Twitter users, upon which we study which properties characterize the influential ones. Our findings show that there are predictive properties associated with the activity level of users and their involvement in communities, but also that writing influential tweets is not a prerequisite for being an influential user.