<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1074">
  <Title>Sydney, July 2006. c(c)2006 Association for Computational Linguistics An Iterative Implicit Feedback Approach to Personalized Search</Title>
  <Section position="3" start_page="4" end_page="585" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Analysis suggests that, while current information retrieval systems, e.g., web search engines, do a good job of retrieving results to satisfy the range of intents people have, they are not so well in discerning individuals' search goals (J. Teevan et al., 2005). Search engines encounter problems such as query ambiguity and results ordered by popularity rather than relevance to the user's individual needs.</Paragraph>
    <Paragraph position="1"> To overcome the above problems, there have been many attempts to improve retrieval accuracy based on personalized information. Relevance Feedback (G. Salton and C. Buckley, 1990) is the main post-query method for automatically improving a system's accuracy of a user's individual need. The technique relies on explicit relevance assessments (i.e. indications of which documents contain relevant information). Relevance feed-back has been proved to be quite effective for improving retrieval accuracy (G. Salton and C.</Paragraph>
    <Paragraph position="2"> Buckley, 1990; J. J. Rocchio, 1971). However, searchers may be unwilling to provide relevance information through explicitly marking relevant documents (M. Beaulieu and S. Jones, 1998).</Paragraph>
    <Paragraph position="3"> Implicit Feedback, in which an IR system unobtrusively monitors search behavior, removes the need for the searcher to explicitly indicate which documents are relevant (M. Morita and Y.</Paragraph>
    <Paragraph position="4"> Shinoda, 1994). The technique uses implicit relevance indications, although not being as accurate as explicit feedback, is proved can be an effective substitute for explicit feedback in interactive information seeking environments (R.</Paragraph>
    <Paragraph position="5"> White et al., 2002). In this paper, we utilize the immediately viewed documents, which are the clicked results in the same query, as one type of implicit feedback information. Research shows that relative preferences derived from immediately viewed documents are reasonably accurate on average (T. Joachims et al., 2005).</Paragraph>
    <Paragraph position="6"> Another type of implicit feedback information that we exploit is users' query logs. Anyone who uses search engines has accumulated lots of click through data, from which we can know what queries were, when queries occurred, and which search results were selected to view. These query logs provide valuable information to capture users' interests and preferences.</Paragraph>
    <Paragraph position="7"> Both types of implicit feedback information above can be utilized to do result re-ranking and query expansion, (J. Teevan et al., 2005; Xuehua Shen. et al., 2005) which are the two general approaches to personalized search. (J. Pitkow et al., 2002) However, to the best of our knowledge, how to exploit these two types of implicit feed-back in a unified way, which not only brings collaboration between query expansion and result re-ranking but also makes the whole system more concise, has so far not been well studied in the previous work. In this paper, we adopt HITS algorithm (J. Kleinberg, 1998), and propose a  HITS-like iterative approach addressing such a problem.</Paragraph>
    <Paragraph position="8"> Our work differs from existing work in several aspects: (1) We propose a HITS-like iterative approach to personalized search, based on which, implicit feedback information, including immediately viewed documents and query logs, can be utilized in a unified way. (2) We implement result re-ranking and query expansion simultaneously and collaboratively triggered by every click. (3) We develop and evaluate a client-side personalized web search agent PAIR, which supports both English and Chinese.</Paragraph>
    <Paragraph position="9"> The remaining of this paper is organized as follows. Section 2 describes our novel approach for personalized search. Section 3 provides the architecture of PAIR system and some specific techniques. Section 4 presents the details of the experiment. Section 5 discusses the previous work related to our approach. Section 6 draws some conclusions of our work.</Paragraph>
  </Section>
class="xml-element"></Paper>