Existing mobility prediction algorithms focus on predicting the next cell or interesting regions such as a home zone. But for position- and movement-based optimization of transmission in a cell such coarse-level mobility prediction is not sufficient. In this paper a learning-based graphical model is introduced which allows a fine-level prediction of the movements and velocities of mobile users inside a cell. We divide the mobile users into different user groups by velocities and learn the path patterns and user type transitional probabilities. Based on this a-priori information a three-step mobility prediction algorithm considering positioning error and future user type is proposed. The simulation result shows a better level of prediction accuracy compared to previous methods.