The 7 th International Workshop on Intelligent Techniques for Web Personalization & Recommender Systems

Recommender systems use a set of reviewers and advice givers with the goal of providing accurate userdependent product predictions. In general, these systems assign weights to different reviewers as a function of their similarity to each user. As products are known to be from different domains, a recommender system also considers product domain information in its predictions. As there are few reviews compared to the number of products, it is often hard to set the similarity-based weights as there is not a large enough subset of reviewers who reviewed the same products. It has then been recently suggested that not considering domains will increase the amount of reviewer data and the overall prediction accuracy in a mediated way. However, clearly, if different reviewers are similar to a user in each product domain, then domain-specific predictions could be superior to mediated ones. In this paper, we consider two advice giver algorithms to provide domain-specific and mediated predictions. We analyze both advice giver algorithms using large real data sets to characterize when each is more accurate for users. We realize that for a considerable number of users, the domain-specific predictions are possible and more accurate. We then contribute an improved general recommender system algorithm that autonomously selects the most accurate mediated or domain-specific advice giver for each user. We validate our analysis and algorithm using real data sets and show the improved predictions for different users. Introduction We model a product recommender system as a set of reviews defined by reviewers, product domains (e.g., DVDs, books, clothes), and advice givers. Users request that the recommender system provide predictions of whether they will like a set of products of their choosing. Users then have the option of providing their own reviews of those products. As the advice giver that makes product predictions receives the user’s actual reviews, it assigns domain-specific weights to the reviewers as a function of the similarity between their reviews and the user’s. The reviewers whose reviews are most similar to the user’s receive higher weight. The advice giver Copyright c © 2009, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. uses the weights of the reviewers and their reviews to guide its predictions with the goal of providing the most accurate predictions. Because reviewers review a relatively small number of products, it is difficult to find enough reviewers with similar reviews to make accurate predictions for every product each user requests. The problem is exacerbated as products are divided into domains so there are fewer product reviews to train the domain-specific weights with. Resolving the data sparsity problem has been the focus of much recommender system work. Although it is widely accepted that domainspecific reviewers result in accurate predictions, it has recently been suggested that a mediated advice giver that combines multiple domains of products and holds only a single set of weights for each, user would help alleviate the data sparsity problem (Berkovsky, Kuflik, and Ricci 2007a). While learning only one set of weights will increase the amount of data to train with, there is an underlying assumption that the reviewers with similar reviews to a user in one product domain (e.g., DVDs) will also have similar reviews to that user in other domains (e.g., books and clothes). While a domain-specific advice giver captures these differences, the mediated advice giver does not. Intuitively, it seems unlikely that for all users there is a set of reviewers with similar reviews in all domains of products. The focus of this work is to understand in real recommender data sets how data sparsity and the user’s reviews affect the weights of the reviewers and the accuracy of the advice givers’ predictions. We first present an overview of how advice givers weigh reviewers and make predictions for users and give examples of weights that affect the two advice givers’ accuracy. We show using data from two large recommender systems that potentially half of the users benefitted from the mediated advice giver while the other half required domain-specific weights. Additionally we find in a third and more sparse recommender data set that both advice givers have equal accuracies when reviewers do not provide reviews for more than one category. We, then, show how accuracy changes for each advice giver as reviewers review products in more categories. Assuming that reviewers do provide reviews in more categories (as found in the first two data sets) and because different users require the two advice givers equally, we contribute two online user-dependent selection algorithms for the recommender system to choose which advice giver