In order to demonstrate privacy threats in social networks we show how to infer user preferences by random walks in a multiple graph representing simultaneously attributes and relationships links. For the approach to scale in a rst phase we reduce the space of attribute values by partition in balanced homogeneous clusters. Following the Deepwalk approach, the random walks are considered as sentences. Hence unsu-pervised learning techniques from natural languages processing can be employed in second phase to deduce semantic similarities of some attributes. We conduct initial experiments on real datasets to evaluate our approach.