User-aware adaptation by subjective metadata and inferred implicit descriptors

Abstract

Universal Media Access faces a wide range of dynamic execution environments with varying networks, terminals, user profiles. In order to handle these highly dynamic contexts, we introduced a formal approach using adaptation agents. Multimedia adaptation agents perceive the context states thanks to observations and carry out adaptation actions. The decision-making principle is based on reinforcement learning and uses Markov Decision Processes. We have already applied this framework to several case studies: adaptive streaming, adaptive content delivery, user profiling etc. Based on our recent work on agent-driven adaptation of multimedia content, this chapter deals with appropriate metadata support for our adaptation framework. Although compatible with MPEG-21, our approach requires changes in the management of metadata. The adaptation agent itself is metadata. Our context descriptors are either observable or only partially observable. The final point is that our decision models naturally handle subjective metadata such as user interest. As a result, we are able to abstract user interest descriptors that are implicitly generated by observing user interactions.