The wide spread of low-cost personal devices equipped with GPS sensors has paved the way towards the creation of customized services based on user mobility habits and able to track and assist users in everyday activities, according to their current location.

In this paper we propose a new approach to extraction and comparison of mobility models, by means of the structure inferred from positioning data. More specifically, we suggest to use concepts and methods borrowed from Algorithmic Learning Theory (ALT) and we formulate mobility models extraction in term of Grammatical Inference (GI), an inductive process able to select the best grammar consistent with the samples and to provide multi-scale generative models. Moreover, we propose a similarity measure by adapting a state-of-the-art metric originally conceived for automata.

A thorough experimental assessment was conducted on the publicly available dataset provided by the Geolife project. Results show how a structural model and similarity metric can provide a better insight on data despite its complexity.