<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1653">
  <Title>Relevance Feedback Models for Recommendation</Title>
  <Section position="3" start_page="0" end_page="449" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Recommender systems (Resnick and Varian, 1997) help users select particular items (e.g, movies, books, music, and TV programs) that match their taste from a large number of choices by providing recommendations. The systems either recommend a set of N items that will be of interest to users (top-N recommendation problem) or predict the degree of users' preference for items (prediction problem).</Paragraph>
    <Paragraph position="1"> For those systems to work, they first have to aggregate users' evaluations of items explicitly or implicitly. Users may explicitly evaluate certain movies as rating five stars to express their preference. These evaluations are used by the systems as explicit ratings (votes) of items or the systems infer the evaluations of items from the behavior of users and use these inferred evaluations as implicit ratings. For example, systems can infer that users may like certain items if the systems learn which books they buy, which articles they read, or which TV programs they watch.</Paragraph>
    <Paragraph position="2"> Collaborative filtering (CF) (Resnick et al., 1994; Breese et al., 1998) and content-based (or adaptive) filtering (CBF) (Allan, 1996; Schapire et al., 1998) are two of the most popular types of algorithms used in recommender systems. A CF system makes recommendations to current (active) users by exploiting their ratings in the database. User-based CF (Resnick et al., 1994; Herlocker et al., 1999) and item-based CF (Sarwar et al., 2001; Karypis, 2001), among other CF algorithms, have been studied extensively. User-based CF first identifies a set of users (neighbors) that are similar to the active user in terms of their rating patterns in the database. It then uses the neighbors' rating patterns to produce recommendations for the active user. On the other hand, item-based CF calculates the similarity between items beforehand and then recommends items that are similar to those preferred by the active user. The performance of item-based CF has been shown to be comparable to or better than that of user-based CF (Sarwar et al., 2001; Karypis, 2001). In contrast to CF, CBF uses the contents (e.g., texts, genres, authors, images, and audio) of items to make recommendations for the active user. Because CF and CBF are complementary, much work has been done to combine them (Basu et al., 1998; Yu et al., 2003; Si and Jin, 2004; Basilico and Hofmann, 2004).</Paragraph>
    <Paragraph position="3"> The approach we took in this study is designed to solve top-N recommendation problems with im- null plicit ratings by using an item-based combination of CF and CBF. The methods described in this paper will be applied to recommending English Wikipedia1 articles based on those articles edited by active users. (This is discussed in Section 3.) We use their editing histories and the contents of their articles to make top-N recommendations. We regard users' editing histories as implicit ratings. That is, if users have edited articles, we consider that they have positive attitudes toward the articles. Those implicit ratings are regarded as positive examples. We do not have negative examples for learning their negative attitudes toward articles. Consequently, handling our application with standard machine learning algorithms that require both positive and negative examples for classification (e.g., support vector machines) is awkward. Our approach is based on the advancement in language modeling approaches to information retrieval (IR) (Croft and Lafferty, 2003) and extends these to incorporate CF. The motivation behind our approach is the analogy between CF and IR, especially between CF and relevance feedback (RF).</Paragraph>
    <Paragraph position="4"> Both CF and RF recommend items based on user preference/relevance judgments. Indeed, RF techniques have been applied to CBF, or adaptive filtering, successfully (Allan, 1996; Schapire et al., 1998). Thus, it is likely that RF can also be applied to CF.</Paragraph>
    <Paragraph position="5"> To apply RF, we first extend the representation of items to combine CF and CBF under the models developed in Section 2. In Section 3, we report our experiments with the models. Future work and conclusion are in Sections 4 and 5.</Paragraph>
  </Section>
class="xml-element"></Paper>