Inferring User Profiles in Online Social Networks Using a Partial Social Graph

Most algorithms for user profile inference in online social networks assume that the full social graph is available for training. This assumption is convenient in a research setting. However, in real-life, the full social graph is generally unavailable or may be very costly to obtain or update. Thus, several of these algorithms may be inapplicable or provide poor accuracy. Moreover, current approaches often do not exploit all the rich information that is available in social networks. In this paper, we address these challenges by proposing an algorithm named PGPI (Partial Graph Profile Inference) to accurately infer user profiles under the constraint of a partial social graph and without training. It is to our knowledge, the first algorithm that let the user control the trade-off between the amount of information accessed from the social graph and the accuracy of predictions. Moreover, it is also designed to use rich information about users such as group memberships, views and likes. An experimental evaluation with 11,247 Facebook user profiles shows that PGPI predicts user profiles more accurately and by accessing a smaller part of the social graph than four state-of-the-art algorithms. Moreover, an interesting result is that profile attributes such as status (student/professor) and gender can be predicted with more than 90% accuracy using PGPI.