Using Implicit Feedback for User Modeling in Internet and Intranet Searchingφ

Powerize Server 1.0, developed by Powerize.com, is a content-based information filtering and retrieval system that presently uses a manually constructed user model known as a search profile. User modeling captures a user’s information needs. A user model can be constructed explicitly by the user or implicitly by exploiting feedback from the user about which documents are relevant. Implicit feedback can be inferred from user behavior without any additional work on the part of the user. The study reported in this paper investigates a way of implementing the implicit feedback technique of user modeling for the Powerize Server 1.0. Previous studies on Internet discussion groups (USENET news) have shown reading time to be a useful source of implicit feedback for predicting a user’s preferences. In this study, we examined: 1) whether reading time is useful for predicting a user’s preferences for academic or professional journal articles, and 2) whether printing behavior adds anything to what we already know from reading time. Two experiments were conducted with undergraduate students using professional articles from the telecommunications and pharmaceutical industries. The results of the experiments showed that reading time could be used to predict the relevancy of documents, although the threshold on reading time required to detect relevant documents would be higher than for USENET news articles. The experiments also showed that printing behavior adds to what can be inferred from reading time. All the documents that were printed in the experiments were relevant, but the reading time for many of these documents was below the mean reading time for all documents read. This result implies that the use of printing behavior with reading time could increase the precision and recall ratios for detecting relevant documents. Suggestions for incorporating the results of the study into the Powerize Server were made in conclusion. This paper also reports detailed technical descriptions of the experiment design, including research problem, experimental system, and data collection. φ The research reported herein was supported in part by the Maryland Industrial Partnerships program and powerize.com