<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-1116">
  <Title>Extraction of User Preferences from a Few Positive Documents</Title>
  <Section position="3" start_page="0" end_page="1" type="intro">
    <SectionTitle>
2 Background
</SectionTitle>
    <Paragraph position="0"> To extract a user's preference from example documents is the same problem as finding their representative vector in linear text classifiers. A variety of algorithms for training linear classifiers have been suggested. Among them, here, we only review two widely used algorithms, Rocchio algorithm and Widrow-Hoff algorithm, for comparing with our method.</Paragraph>
    <Paragraph position="1"> The Rocchio algorithm (David et al., 1996) is a batch algorithm. So, it produces a new weight vector w from an existing weight vector</Paragraph>
    <Paragraph position="3"> lyzing the entire set of training data at once. The j th' component of w is :</Paragraph>
    <Paragraph position="5"> x and n is the number of training documents. C is the set of positive training documents, and c n is the number of positive training documents. The parameter ba, and g control the relative impact of the original weight vector, the positive examples, and the negative examples, respectively. However, in our experiments, a = 0, b =1, and g = 0 because only positive examples are given in our application. Neither original weight vector nor negative examples is given. The Widrow-Hoff algorithm (David et al., 1996) is an online algorithm where one training example is presented at a time. It updates its current weight vector based on the example and then discards the example, retaining only the new weight vector. A new weight vector w  is computed from an old</Paragraph>
    <Paragraph position="7"> x is in the set of positive or relevant training documents, otherwise 0. In our application, i y is always 1 because we deal with only positive examples. The initial weight vector w  is typically set to zero vector, w</Paragraph>
    <Paragraph position="9"> where, e is the learning rate which controls how quickly the weight vector w is allowed to change and ii xw * is the cosine value of the two vectors.</Paragraph>
  </Section>
class="xml-element"></Paper>