Maschinelles Lernen Bayes'scher Netze für benutzeradaptive Systeme

This thesis addresses machine learning techniques for Bayesian networks in the context of useradaptive systems. On the basis of the definition of a Bayesian network and the discussion of related methods and extensions of this framework, an overview of current research on the application of Bayesian networks within user-adaptive systems is presented. So far, there have been no major efforts to develop Bayesian network learning algorithms that are especially well suited to dealing with the demands of the user modeling context, or that are able to exploit the specific characteristics of this context. The general formulation of the machine learning problem is transferred to the context of useradaptive systems. Several crucial criteria are identified that have to be addressed adequately during the planning phase to ensure a successful application of machine learning techniques in a useradaptive system. The methods that are developed in this thesis, along with already existing standard learning methods, are integrated into a general framework that can be instantiated according to the demands of the domain to be modeled. This generic framework describes the relationships between available data, prior knowledge, offline acquisition of Bayesian network user models, and the models’ online adaptation. The purpose of this framework is to address the identified criteria for the successful application of machine learning methods in user-adaptive systems in the case of Bayesian networks. The system’s developer can choose an adequate selection from a repertoire of existing and new methods that are presented in this thesis. This selection can be arranged within the generic framework, thereby yielding a specific instance of the framework. Typically, only a subset of the whole range of possibilities has to be implemented in order for the demands of the domain under consideration to be satisfied. The following concrete contributions are made by this thesis: • Identification of crucial criteria for a successful application of machine learning methods in the context of user-adaptive systems and the discussion of solutions for the case of Bayesian network user models • Integration of existing and newly developed methods into a generic framework for learning Bayesian networks for user-adaptive systems • Development of several machine learning algorithms for Bayesian networks with focus on the demands and characteristics of the user modeling context: