Exploring Versus Exploiting when Learning User Models for Text Recommendation

• Authors: Marko Balabanović
• Venue: User Modeling and User-Adapted Interaction
• Year: 1998

The text recommendation task involves delivering sets of documents tousers on the basis of user models. These models are improved overtime, given feedback on the delivered documents. When selectingdocuments to recommend, a system faces an instance of theexploration/exploitation tradeoff: whether to deliver documents aboutwhich there is little certainty, or those which are known to match theuser model learned so far. In this paper, a simulation is constructedto investigate the effects of this tradeoff on the rate of learninguser models, and the resulting compositions of the sets of recommendeddocuments, in particular World-Wide Web pages. Document selectionstrategies are developed which correspond to different points alongthe tradeoff. Using an exploitative strategy, our results show thatsimple preference functions can successfully be learned using avector-space representation of a user model in conjunction with agradient descent algorithm, but that increasingly complex preferencefunctions lead to a slowing down of the learning process. Exploratorystrategies are shown to increase the rate of user model acquisition atthe expense of presenting users with suboptimal recommendations; inaddition they adapt to user preference changes more rapidly thanexploitative strategies. These simulated tests suggest animplementation for a simple control that is exposed to users, allowingthem to vary a system‘s document selection behavior depending onindividual circumstances.