Computer-based training systems are of growing importance nowadays. Demand for remote skill and knowledge validating solutions is higher than ever before. Serious games provide safe and replayable conditions for teaching and practice. The main issue with education founded on serious games paradigm is the evaluation of the players. Virtual environments allow for various activities that are hard to grasp and quantify automatically. Gameplay log files are usually filled with an excess of trivial activities that don’t provide vital information about user’s skills or knowledge in the relevant domain. In this paper, the authors propose a novel method for classification and interpretation of user activities in computer-based training systems. The innovation of the presented idea lies in the combination of layered graphs data structures, and Bayesian inference about the user’s skills and knowledge.