Predicting complex user behavior from CDR based social networks

Abstract Call Detail Record (CDR) datasets provide enough information about personal interactions of cell phone service customers to enable building detailed social networks. We take one such dataset and create a realistic social network to predict which customer will default on payments for the phone services, a complex behavior combining social, economic, and legal considerations. After extracting a large feature set from this network, we find that each feature poorly correlates with the default status. Hence, we develop a sophisticated model to enable reliable predictions. Our main contribution is a methodology for building complex behavior models from very large sets of diverse features and using different methods to choose those features that perform best for the final model. This approach enables us to identify the most efficient features for our problem which, unexpectedly, are based on the number of unique users with whom the given user communicates around the Christmas and New Year’s Eve holidays. In general, features based on the number of close ties maintained by a user perform better than others. Our resulting models significantly outperform the methods currently published in the literature. The paper contributes also a systematic analysis of properties of the network derived from CDR.