Efficient prediction model management in mobile systems

Abstract

With the advent of affordable mobile devices such as smartphones and tablets, the vision of Pervasive Computing has made a big step closer to becoming reality. In order to become truly ubiquitous and seamlessly integrate into everyday life, the design of context-aware applications is essential. Using contextual information obtained for example from the device's sensors such as motion sensors and gps receiver, context-aware applications can adapt their behavior depending on the environment the user is in. In some scenarios, context aware applications can also benefit from knowledge about future contexts. This necessitates the use of a context prediction model. We examine a social network scenario where in addition, the context in question is originally being acquired on another user's device. In this scenario, the prediction model could for example be used to predict the next location or activity of a friend. Prior to that, the prediction model needs to be distributed to and stored on the mobile device running the application. Both high transfer cost and limited space make it imperative to produce small prediction models which still predict the context considerably well. In this thesis, we examined methods to compress Markov-based prediction models of higher order in a lossless and lossy fashion and evaluated these methods on real world and generated data. Our evaluation showed clearly that the compression mechanisms introduced can be successfully applied to significantly reduce the size of the prediction models with only a minor impact on prediction performance.