<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3801">
  <Title>A Graphical Framework for Contextual Search and Name Disambiguation in Email</Title>
  <Section position="6" start_page="6" end_page="7" type="relat">
    <SectionTitle>
6 Related Work
</SectionTitle>
    <Paragraph position="0"> As noted above, the similarity measure we use is based on graph-walk techniques which have been adopted by many other researchers for several different tasks. In the information retrieval community, infinite graph walks are prevalent for determining document centrality (e.g., (Page et al., 1998; Diligenti et al., 2005; Kurland and Lee, 2005)). A related venue of research is of spreading activation over semantic or association networks, where the underlying idea is to propagate activation from source nodes via weighted links through the network (Berger et al., 2004; Salton and Buckley, 1988).</Paragraph>
    <Paragraph position="1"> The idea of representing structured data as a graph is widespread in the data mining community, which is mostly concerned with relational or semi-structured data. Recently, the idea of PageRank  has been applied to keyword search in structured databases (Balmin et al., 2004). Analysis of inter-object relationships has been suggested for entity disambiguation for entities in a graph (Kalashnikov et al., 2005), where edges are unlabelled. It has been suggested to model similarity between objects in relational data in terms of structural-context similarity (Jeh and Widom, 2002).</Paragraph>
    <Paragraph position="2"> We propose the use of learned re-ranking schemes to improve performance of a lazy graph walk.</Paragraph>
    <Paragraph position="3"> Earlier authors have considered instead using hill-climbing approaches to adjust the parameters of a graph-walk (Diligenti et al., 2005). We have not compared directly with such approaches; preliminary experiments suggest that the performance gain of such methods is limited, due to their inability to exploit the global features we used here6. Related research explores random walks for semi supervised learning (Zhu et al., 2003; Zhou et al., 2005).</Paragraph>
    <Paragraph position="4"> The task of person disambiguation has been studied in the field of social networks (e.g., (Malin et al., 2005)). In particular, it has been suggested to perform name disambiguation in email using traffic information, as derived from the email headers (Diehl et al., 2006). Our approach differs in that it allows integration of email content and a timeline in  framework. In addition, we incorporate learning to tune the system parameters automatically.</Paragraph>
  </Section>
class="xml-element"></Paper>