COD: Iterative Utility Elicitation for Diversified Composite Recommendations Khalid Alodhaibi George Mason University 4400 University Drive 4A4 Fairfax, VA 22030 kalodhai@gmu.edu Alexander Brodsky George Mason University 4400 University Drive 4A4 Fairfax, VA 22030 brodsky@gmu.edu George A. Mihaila IBM T.J. Watson Res. Ctr. 19 Skyline Drive Hawthorne, NY, USA mihaila@us.ibm.com Abstract This paper studies and proposes methods for provid- ing recommendations on composite bundles of products and services that are dynamically defined using database views extended with decision optimization based on mathematical programming. A framework is proposed for finding a diverse recommendation set when no prior knowledge on user preference is given. To support this framework, a method is developed for utility function elicitation, which is based on iteratively refining a set of axes in the n-dimensional utility space. The notion of a diverse recommendation set is refined and formal- ized by partitioning the recommendation space into lay- ers that correspond to their distance to the maximal util- ity. In each layer, the method selects recommendations that maximize each dimension of the utility space. A pre- liminary experimental study is conducted, which shows that the proposed framework significantly outperforms a popular commercial system in terms of precision and re- call. 1 Introduction Recommender systems are increasingly used to help with selection of diverse products and services over the Internet. This paper focuses on recommending compos- ite services and products and eliciting user preferences. Most of todays recommender systems recommend only atomic products or services. Complex recommendation models involving composite alternatives, such as prod- uct configurations and service packages, are rarely ad- dressed. In addition, the majority of recommender sys- tems rely on a single ranking or utility score, whereas, in many applications, there are multiple criteria that need to be taken into account, such as price, quality and en- joyment. Recently, multi-criteria ranking has been explored in recommendation set retrieval [2,15]. These methods choose a set of alternatives based on a distance measure calculated for each of the multiple criteria. Multi-criteria ranking can help provide a balance between diversity and optimality. However, most recommender systems limit recommendations to those that are relevant to users re- quests. Therefore, their recommendations are often simi- lar to each other and do not provide enough diversity. Di- versity is important because it helps users become aware of choices they may not have thought of. With the recent surge in collaborative similarity-based recommenders, such as Amazon.com, a number of multi-criteria ranking methods have been proposed. Of significant importance to this research is work suggest- ing the importance of diversity sensitive recommenda- tion sets. The work presented in [2,12] details several algorithms for selecting diverse recommendation alter- natives based on the similarity of individual attributes. The work done by Linden, et al [9] also suggests a di- verse ranking algorithm. Zhang and Hurley [23] used a similar approach with respect to calculating diversity; however, their similarity measure of recommendations was based on a set rather than individual recommenda- tions. For example, a recommendation with low simi- larity to the target might make it to the final list because the similarity score of the set it belongs to, is above a threshold. Furthermore, most existing recommender systems are designed for a single target domain and do not provide a general framework for the development of recommender systems. Finally, many recommender systems are intru- sive and require explicit and significant feedback from the user [1]. Feedback will continue to be a primary fac- tor in the recommender system concept; however, the next generation of recommender systems might want to extract information from users implicitly. An example might be how long the user spends reading a specific document to infer how much the user liked the docu- ment, consequently, giving it a higher rating. There are several approaches for eliciting utility func- 1 Proceedings of the 43rd Hawaii International Conference on System Sciences - 2010 978-0-7695-3869-3/10 $26.00 © 2010 IEEE