This paper focuses on predicting drop-outs and school failures when student data has been enriched with data derived from students social behaviour. These data describe social dependencies gathered from e-mail and discussion board conversations, among other sources. We describe an extraction of new features from both student data and behaviour data represented by a social graph which we construct. Then we introduce a novel method for learning a classifier for student failure prediction that employs cost-sensitive learning to lower the number of incorrectly classified unsuccessful students. We show that the use of social behaviour data results in significant increase of the prediction accuracy. (Contains 2 figures, 7 tables, and 1 footnote.) [This work has been partially supported by the Faculty of Informatics, Masaryk University. For the complete proceedings, "Proceedings of the International Conference on Educational Data Mining (EDM) (5th, Chania, Greece, June 19-21, 2012)," see ED537074.]