@phdthesis{601, keywords = {personalization, Recommender systems, collaborative filtering, profil model, Context, Information}, author = {Erich Christoph Gstrein}, title = {Adaptive personalization : a multi view personalization approach incorporating contextual information}, abstract = {

Information overload has become a significant problem causing losses in economic relevant dimensions. According to a study of Basex (http://www.basex.com) published in 12/2008 only the U.S. economy loses at least $900 billion per year due to lowered employee productivity and reduced innovation. Personalization systems, generating individual suggestions based on user models, are seen as one major concept for solving this urgent problem of information overload. However, providing an appropriate personal assistance requires a diligent usage of the available information.

In this thesis we propose a personalization approach - called Adaptive Personalization - utilizing the available, context specific information in a new and efficient way. In addition to other approaches tackling the incorporation of contextual information, we only rely on minimal data available in standard recommendation scenarios. We will show how the knowledge about the origin of user feedback - called context of origin - can be used to construct an enhanced multi-view profile model, also incorporating long- and short-term preference aspects as well as neglect. For solving the k-nearest neighbors problem in a high dimensional search space, stretched by the attributes of these profiles, a new index structure - called D2[D hoch 2]-Tree - will be presented supporting the dynamic change of (contextual) distance functions.

Furthermore, we introduce a new collaborative filtering algorithm where techniques borrowed from information retrieval and association rule mining where used to improve prediction quality, outperforming established approaches especially in contexts with minimal information.

Additionally, a robust and flexible architecture is presented suitable for large scale, real world application scenarios together with a proposal for a procedure model concerning system design. While the effectiveness of our profile model was determined on the basis of real world data, the collaborative filtering algorithm was evaluated using popular test data sets and metrics. Furthermore, detailed cost analyzes concerning the different index operations are provided for the D2[D hoch 2]-Tree.

Due to this new and efficient utilization of the available information, its light-weight profile model, the efficient algorithms and the flexible architecture, the Adaptive Personalization approach is a useful and valuable contribution for solving the problem of information overload.

}, year = {2009}, journal = {Institute of Visual Computing and Human-Centered Technology}, pages = {210}, publisher = {TU Wien}, address = {Vienna}, }