PEBL: Web Page Classification without Negative Examples

1
PEBL Web Page Classification without Negative
Examples

• Hwanjo Yu, Jiawei Han, and Kevin Chen-Chuan
• IEEE Transactions on Knowledge and Data
  Engineering, Vol. 16, No. 1, 2004
• Presented by Chirayu Wongchokprasitti

2
Introduction

• Web page classification is one of the main
  techniques for Web mining
• Constructing a classifier requires positive and
  negative training examples
• Cautious to avoid bias and laborious to collect
  negative training examples

3
Typical Learning Framework
4
Positive Example Base Learning (PEBL) Framework

• Learn from positive data and unlabeled data
• Unlabeled data indicates random samples of the
  universal set
• Apply the Mapping-Convergence (M-C) Algorithm

5
Mapping-Convergence (M-C) Algorithm

• Divide into 2 stages
• Mapping stage
• Use any classifier that does not generate false
  negatives
• They chose 1-DNF ( monotone Disjunctive Normal
  Form)
• Convergence stage
• For maximizing margin
• They chose SVM (Support Vector Machine)

6
Mapping Stage

• Use a weak classifier to draw an initial
  approximation of strong negative data.
• First, Identify strong positive features from
  positive and unlabeled data by checking the
  frequency of those features.
• If feature frequency in positive data is larger
  than one in the universal data, it is a strong
  positive
• Filter out any possible positive, leaving only
  strong negatives.

7
Convergence Stage

• Use SVM to scope down the class boundary
• Iterate SVM for certain times to extract negative
  data from unlabeled data
• The boundary will converge into the true boundary.

8
Support Vector Machines
Visualization of a Support Vector Machine
9
Convergence of SVM
10
Data Flow Diagram
11
Experimental Results

• Report the result with precision-recall breakeven
  point (P-R)
• Experiment 1 the Internet
• Use DMOZ as the universal set
• Experiment 2 University CS department
• Use WebKB data set
• Mixture Models

12
Experiment 1
13
Experiment 2
14
Mixture Models
15
Summary and Conclusions

• PEBL framework eliminates the need for manually
  collecting negative training examples
• The Mapping-Convergence (M-C) algorithm achieves
  classification accuracy as high as that of
  traditional SVM
• PEBL needs faster training time