How Are You Related? Predicting the Type of a Social Relationship Using Call Graph Data

Social relationships defined by phone calls made between people can be grouped into various relationship types or categories, such as family members, co-workers, etc. We propose and evaluate a method that predicts the "relationship type" between a pair of mobile phone subscribers using features that abstract their communication behavior and social network patterns. Our dataset consists of call detail records of a major wireless carrier sampled from four demographically diverse regions, from which we built a directed social graph, with over 200,000 vertices and 400,000 edges. Using account and subscription plan information, we labeled each edge in the graph as one of the following four relationships: family, co-worker, customer and service. Our analysis of the dataset shows that these four relationship types exhibit distinct communication behavior patterns and generate characteristic topological features on the social network surrounding the pairs. For instance, subscriber pairs with a family relationship generate high average number of calls, have low call duration, call more frequently and share more mutual contacts than pairs with a service or co-worker relationship. Using a set of features that abstract these characteristics and the Random Forest supervised machine learning classifier, we demonstrate that it is possible to predict the relationship type between a subscriber pair with an accuracy of 87%.