User model induction for intelligent information access

The explosive growth of information available on the Internet has created a clear need for novel methods that help users locate relevant information quickly and with minimal effort. The central argument of this dissertation is that learning about users' multiple and potentially changing interests calls for algorithms specifically designed for this purpose. Guided by this principle, I introduce the Adaptive Information Server ( AIS), a client-server framework for domain-independent adaptive information access. I describe two applications that use AIS to learn about users' interests in daily news stories: one system operates on the World Wide Web, the other is geared towards wireless information access. The description and evaluation of the underlying recommendation algorithms form the core of the dissertation.;First, I describe a content-based learning algorithm designed to learn about users' multiple and frequently changing interests. The key to the algorithm's performance lies in its multi-strategy design: it learns separate models of users' short-term and long-term interests. An empirical evaluation shows that the combination of both models performs better than each individual model alone. In addition, the algorithm maintains a model of information the user is likely to know, so that the presentation of redundant content can be avoided. Second, I show how the described content-based algorithm can be extended with a collaborative filtering component. In particular, I cast collaborative filtering as a learning task, and present a novel algorithm that uses the Singular Value Decomposition to derive a low-dimensional data representation that forms the basis of an efficient and accurate approach to collaborative filtering. Empirical results demonstrate that the resulting approach outperforms previously proposed algorithms, and that combining content-based and collaborative techniques leads to overall performance improvements.;The dissertation concludes with a description of two empirical studies that evaluate the utility of adaptive information access from a user perspective. These studies show that the adaptive presentation of personalized content simplifies access to relevant information, and that the observed performance can be achieved without requiring any extra work from the user.