VirtualWorlds (VW) have gained popularity in the last years in domains like training or education mainly due to their highly immersive and interactive 3D characteristics. In these platforms, the user (represented by an avatar) can move and interact in an artificial world with a high degree of freedom. They can talk, chat, build and design objects, program and compile their own developed programs, or move (flying, teleporting, walking or running) to different parts of the world. Although these environments provide an interesting working place for students and educators, VW platforms (such as OpenCobalt or OpenSim amongst others) rarely provide mechanisms to facilitate the automatic (or semi-automatic) behaviour analysis of users interactions. Using a VW platform called VirtUAM, the information extracted from different experiments are used to analyse and define students communities based on their behaviour. To define the individual student behaviour, different characteristics are extracted from the system, such as the avatar position (in form of GPS coordinates) and the set of actions (interactions) performed by students within the VW. Later this information is used to automatically detect behavioural patterns. This paper shows how this information can be used to group students in different communities based on their behaviour. Experimental results show how community identification can be successfully perform using K-Means algorithm and Normalized Compression Distance. Resulting communities contains users working in near places or with similar behaviours inside the virtual world.