Résumé
Gradual patterns have been studied for many years as they contain precious information. They have been integrated in many expert systems and rule-based systems, for instance to reason on knowledge such as {\\em ''the greater the number of turns, the greater the number of car crashes\"}. In many cases, this knowledge has been considered as a rule {\\em ''the greater the number of turns $\\rightarrow$ the greater the number of car crashes\"} Historically, works have thus been focused on the representation of such rules, studying how implication could be defined, especially fuzzy implication. These rules were defined by experts who were in charge to describe the systems they were working on in order to turn them to operate automatically. More recently, approaches have been proposed in order to mine databases for automatically discovering such knowledge. Several approaches have been studied, the main scientific topics being: how to determine what is an relevant gradual pattern, and how to discover them as efficiently as possible (in terms of both memory and CPU usage). However, in some cases, end-users are not interested in raw level knowledge, and are rather interested in trends. Moreover, it may be the case that no relevant pattern can be discovered at a low level of granularity (e.g. city), whereas some can be discovered at a higher level (e.g. county). In this paper, we thus extend gradual pattern approaches in order to consider multiple level gradual patterns. For this purpose, we consider two aggregation policies, namely horizontal and vertical.