<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-2408">
  <Title>Modeling Category Structures with a Kernel Function</Title>
  <Section position="8" start_page="64" end_page="64" type="concl">
    <SectionTitle>
7 Conclusion
</SectionTitle>
    <Paragraph position="0"> We proposed a TOP kernel based on separating hyperplanes. The proposed kernel is created from one-dimensional Gaussians along the normal directions of the hyperplanes. We showed that the computational advantage that the proposed kernel has is shared by a more general class of models. We empirically showed that the proposed kernel outperforms the linear kernel in text categorization. null Although the superiority of the proposed method to the linear kernel was shown, the proposed method has to be further investigated. Firstly, for large data sizes (namely 7000 and 8000), the proposed method was not significantly better than the linear kernel. The effectiveness of the proposed method should be confirmed by more experiments and theoretical analysis. Secondly, we have to compare the proposed method with other kernels in order to check the effectiveness of the kernel function consisting of one-dimensional Gaussians normal to the hyperplanes. The use of Gaussians is open to argument, because their symmetric form is somewhat against our 4If the computational time required for feature extraction is included, the HP-TOP kernel cannot be faster than the linear kernel.</Paragraph>
    <Paragraph position="1"> intuition.</Paragraph>
    <Paragraph position="2"> This model can be extended to incorporate unlabeled examples, for example, using the EM algorithm. In that sense, the combination of PLSI and the semi-supervised EM algorithm is also one promising model. When the category structure of the negative examples is not given, the proposed method is not applicable. We should investigate whether unsupervised clustering can substitute for the category structure.</Paragraph>
  </Section>
class="xml-element"></Paper>