This paper presents a set of methods for the analysis of user activity and data preparation for the music recommender by the example of “Odnoklassniki” social network. The history of actions is being analyzed in multiple dimensions in order to find a number of collaborative and temporal correlations as well as to make the overall rankings. The results of the analysis are being exported in a form of a taste graph which is then used to generate on-line music recommendations. The taste graph displays relations between different entities connected with music (users, tracks, artists, etc.) and consists of the following main parts: user preferences, track similarities, artists’ similarities, artists’ works and demography profiles.