HYRIWYG: Leveraging Personalization to Elicit Honest Recommendations Ana Cristina Bicharra Garcia Instituto de Computação Universidade Federal Fluminense Niterói, RJ, Brazil 24210-240 +51 (21) 2629 5675 bicharra@stanford.edu Martin Ekström Stanford University CIFE, Terman Engineering Center Stanford, CA 94305 +1 (650) 725 0406 mekstrom@stanfordalumni Hans Björnsson Stanford University CIFE, Terman Engineering Center Stanford, CA 94305 +1 (650) 723 6486 hansbj@stanford.edu ABSTRACT This paper presents HYRIWYG (How You Rate Influences What You Get), a reputation system applicable to Internet Recommendation Systems (RS). The novelty lies in the incentive mechanism that encourages evaluators to volunteer their true opinion. Honesty is encouraged because rewards are indexed by the quality of the RS’s suggestions. Categories and Subject Descriptors H.3.3 [Information Storage and Retrieval]: Relevance Feedback. General Terms Management, Economics, Reliability Keywords Recommender Systems, Reputation, Trust, Incentives, Personalization, Game Theory. 1. INTRODUCTION Internet users often draw on Recommendation Systems (RS) [1] as a source of valuable information. Recommendations are based on users’ purchasing behavior, demographics and evaluations provided by peer consumers, which causes potential incentive problems. First, since RSs are public goods, there is an intrinsic risk of free rider behavior. Long and complex feedback forms aggravate this problem. Second, people usually provide evaluations according to hidden agendas, which compromise the quality of the RS. To mitigate both of these problems, this paper proposes an incentive mechanism, HYRIWYG (How You Rate Influences What You Get) that encourages evaluators to volunteer their true opinion. Honesty is encouraged because rewards are indexed by the quality of the RS’s suggestions. Thus, dishonest evaluators end up with unwanted products. An HYRIWYG system could be quite useful for a movie rental site, for example, by providing coupons for recommended movies. 2. HYRIWYG: H OW Y OU R ATE I NFLUENCES W HAT Y OU G ET HYRIWYG is a new reputation mechanism [2] to provide users with the right incentive to volunteer truthful evaluations on which the Recommendation Systems can base their suggestions. Consequently, it is an indirect way to enhance the RS’s credibility and also to tune the RS inference mechanism. Although the long- run benefits of providing evaluations to build the knowledge base and to tune the RS may be clear, users do not perceive this gain immediately. HYRIWYG ’s incentive is to pay users up front, but with a form of compensation that is as good as the quality of the RS being evaluated. Consequently, HYRYWIG makes user and RS enter into a virtuous mode. Let us assume a set of evaluators I= 1, 2..N. Each evaluator i has a type θ i ; i.e., agent i's preference function. The outcome vector is x = (K, τ 1 , … τ i ), where K is the aggregate product evaluation and τ is the incentive provided for each one of the N evaluators. The values for K vary depending on the RS evaluation model, such as Bad/Good or 1-5 stars and so on. Agent i wants to maximize his utility function (u i ). Consequently, if he has more work to do, he must be rewarded for that extra work. In the absence of incentives, a RS requires people to act altruistically. The incentive, or the social function, plays a critical role in bringing the benefit from tuning the RS to the moment a person is providing the recommendation. Sometimes, people can recognize the indirect benefit of acting truthfully when providing their opinion about a product because they are frequent users of the RS suggestions, such as in movie rental sites [3]. We propose a social function that provides incentives for each individual to tune the RS for his profile. As shown in the compensation function below (Eq. 1), the incentive is a constant C – corresponding to points, coupons, savings or any other reward according to the product’s marketing strategy. C is adjusted by the contribution of agent i towards the tuning of the RS. Compensation Function User i’s compensation for submitting an individual recommendation, thus, becomes: τ i = C ∗ (1 + α ∗ | v’(RS(θ i )) – v i |), (1) where Copyright is held by the author/owner(s). EC’04, May 17–20, 2004, New York, New York, USA. ACM 1-58113-711-0/04/0005. 232