The World Wide Web provides access to a wealth of information and services to a huge and heterogeneous user population on a global scale. One important and successful design mechanism in dealing with this diversity of users is to personalize Web sites and services, i.e. to customize system content, characteristics, or appearance with respect to a specific user. Each system independently builds up user profiles and uses this information to personalize the service offering. Such isolated approaches have two major drawbacks: firstly, investments of users in personalizing a system either through explicit provision of information or through long and regular use are not transferable to other systems. Secondly, users have little or no control over the information that defines their profile, since user data are deeply buried in personalization engines running on the server side. Cross system personalization (CSP) (Mehta, Niederee, & Stewart, 2005) allows for sharing information across different information systems in a user-centric way and can overcome the aforementioned problems. Information about users, which is originally scattered across multiple systems, is combined to obtain maximum leverage and reuse of information. Our initial approaches to cross system personalization relied on each user having a unified profile which different systems can understand. The unified profile contains facets modeling aspects of a multidimensional user which is stored inside a "Context Passport" that the user carries along in his/her journey across information space. The user’s Context Passport is presented to a system, which can then understand the context in which the user wants to use the system. The basis of ’understanding’ in this approach is of a semantic nature, i.e. the semantics of the facets and dimensions of the unified profile are known, so that the latter can be aligned with the profiles maintained internally at a specific site. The results of the personalization process are then transfered back to the user’s Context Passport via a protocol understood by both parties. The main challenge in this approach is to establish some common and globally accepted vocabulary and to create a standard every system will comply with. Machine Learning techniques provide an alternative approach to enable CSP without the need of accepted semantic standards or ontologies. The key idea is that one can try to learn dependencies between profiles maintained within one system and profiles maintained within a second system based on data provided by users who use both systems and who are willing to share their profiles across systems – which we assume is in the interest of the user. Here, instead of requiring a common semantic framework, it is only required that a sufficient number of users cross between systems and that there is enough regularity among users that one can learn within a user population, a fact that is commonly exploited in collaborative filtering. In this thesis, we aim to provide a principled approach towards achieving cross system personalization. We describe both semantic and learning approaches, with a stronger emphasis on the learning approach. We also investigate the privacy and scalability aspects of CSP and provide solutions to these problems. Finally, we also explore in detail the aspect of robustness in recommender systems. We motivate several approaches for robustifying collaborative filtering and provide the best performing algorithm for detecting malicious attacks reported so far.