This paper proposes a new layer for a novel IPTV service that is focused on the distribution of personalized multimedia contents over IP networks based on the concept of content-zapping, in contrast to traditional channel-zapping. Our system aims to aggregate continuously the available multimedia contents in its providers and suggest them efficiently to its endusers. Thus, the system will allow users to receive a multimedia streaming that is automatically and specifically composed for them. At the same time, the service will dynamically classify its users, multimedia contents, multimedia providers and other crucial entities. This classification is dynamically built and gradually refined based on the interactions between all the system entities. A specific content must be suggested to one or more users not only based on the content's and user's profiles but also based on dynamic context parameters, such as the user location, the current date/time, the user client device, the user environment, etc. In order to synthesize and save all the knowledge, a classification methodology, learning algorithms and recommendation techniques were defined. All the proposed mechanisms were articulated in a flexible and robust architecture. A client prototype was also developed and tested together with the proposed system.