Understanding urban mobility is a fundamental question for institutional organizations (transport authorities, city halls) and it involves many different fields like social sciences, urbanism or geography. With the increasing number of probes tracking human locations, like RFID pass for urban transportation, road sensors, CCTV systems or cell phones, mobility data are exponentially growing. Mining the activity logs in order to model and characterize efficiently our mobility patterns is a challenging task involving large scale noisy datasets. In this article, we present a robust approach to characterize activity patterns from the activity logs of a urban transportation network. Our study focuses on the Paris subway network. Our dataset includes more than 80 millions travels made by 600 k users. The proposed approach is based on a multi-scale representation of the user activities, extracted by a nonnegative matrix factorization algorithm (NMF). NMF is used to learn dictionaries of usages that can be exploited in order to characterize user mobility and station patterns. The relevance of the extracted dictionaries is then assessed by using them to cluster users and stations. This analysis shows that public transportation usage patterns are tightly linked to sociological patterns. We compare our approach with a k-means baseline that does not take into account user information and demonstrate the interest of characterizing user profiles to obtain better representations of stations.