Beyond Search: Proactive Document Recommendation With Bayesian Graphical Models

Yi Zhang
University of California, Santa Cruz

Abstract

When a user has a long term information need, search engines may not be the best solution. An alternative is a personal filtering system, a proactive document recommendation system that learns from the user and pushes relevant documents to the user in a dynamic environment. To develop an intelligent personal filtering system, the biggest challenge is to adaptively learn user profiles from limited user supervision. If we ask a human agent to solve this problem, he/she may consult with domain experts, borrow information from other users, integrate the content of each document with other forms of evidence to infer about the user's information needs, and do active learning by carefully picking the right questions to ask the user so that the answer can provide the most valuable information. Motivated by this, I will present a set of solutions that enable a filtering system behave similar to a human agent based on Bayesian Theory and Graphical Models.