Learning and scalability in personalized information retrieval and filtering

Abstract: This licentiate thesis is composed of three papers on personalized information retrieval and filtering. The first paper deals with personalized information retrieval. A method is presented that learns from user feedback in a long-term fashion. An experimental evaluation of the method on a standard test collection demonstrates highly improved precision and recall for novel user queries. In the second paper, a technique for large-scale personalized collaborative filtering is developed. The technique is based on inverted files, similar to algorithms used in text retrieval. Experimental results show that the inverted file search strategy is many times faster than the in-memory base line method. Further improvements are also demonstrated by the use of early termination heuristics. The third paper describes the architecture and implementation of a system for personalized information filtering. The filtering functionality is general, and may be used for several tasks such as text classification and collaborative filtering. The system may be used in a stand-alone application, or in a client-server environment. Early versions of the system have already been used successfully in a number of research projects.