In order to adapt to the environment of the user, devices have to be able to deduce the user?s goals and information needs. In mobile environments, the goals and information needs of the user potentially depend on the user?s situation. Existing work on context-dependent user modeling has mainly focused on specific application domains, most notably location-based services such as tourist guides, or on technological enablers. What is currently lacking is an understanding of when and why different personalization techniques work or fail. In this paper, we compare different classification algorithms on data collected from a mobile application. Our results show that methods that are able to learn tree-structured dependencies seem good candidates for personalization due to (i) the inherent hierarchical nature of context information and (ii) the fast running time of the algorithms. We also suggest two future research issues: (1) obtaining a better understanding of the nature of dependencies in contextual data, and (2) using collaborative user modeling techniques to improve the predictive power of user models.