Let's imagine a system that can recommend the kind of mu-sic (among other application domains) you like to listen when you are at work, without having to know your location, IP address or even to ask your current mood. In this paper, we bring this dream closer by proposing a model that can automatically understand the user's current context. This model, called DANCE, analyzes the attributes of the items in your recent history and monitors the relative diversity brought by your consultations over time. We validated our approach with a music corpus of 100 users and a global history of 204,758 plays.