Using Web Structure for Classifying and Describing Web Pages

1
Using Web Structure for Classifying and
Describing Web Pages

• Eric J. Glover1, Kostas Tsioutsiouliklis1,2,
• Steve Lawrence1, David M. Pennock1, Gary W.
  Flake1
• International World Wide Web Conference, 2002
• Presented by Zaihan Yang
• CSE Web Mining

2
Introduction

• Aim
• Classification of web pages
• Description of web pages (to name clusters of web
  pages)
• Using Web Structure
• Extracting patterns from hyperlinks in the web.
• HyperLink
• The destination page
• Associated anchortext describing link

3

• Typical Text-based classification
• To utilize the words (or phrases) of a target
  document, considering the most significant
  features.
• Not Effective.
• E.g.
• The home page of General Motors (www.gm.com)
  does not state that they are a car company.

• Full text
• Anchortext
• Extended-anchortext
• A combination

4
Virtual Document

• A virtual document is
• A collection of anchortexts or extended
  anchortexts from links pointing to the target
  document.
• Anchortext
• The words occurring inside of a link
• Extended anchortext
• The set of rendered words occurring up to 25
  words before and after an associated link (as
  well as the anchortext itself).

5

• Main Method
• Full-text classifier
• Virtual documents classifier
• Two Improvement methods
• Name a cluster
• Main Procedure

6

• Datasets
• Positive a set of web pages downloaded from
  various Yahoo! Categories.
• Negative Random documents from outside Yahoo!
• WebKB dataset
• Features
• All words and two or three word phrases
• i.e. My favorite game is scrabble.
• Possible features
• My, my favorite, my favorite game, favorite,
  favorite game, etc.

7
Dimensionality reduction

• To remove useless features.
• Two step process
• First, remove all features that do not occur in a
  specified percentage of documents.
• i.e. (Af/A lt T) and (Bf/B lt T-)
• A the set of positive examples.
• B the set of negative examples.
• Af documents in A that contain feature f.
• Bf documents in B that contain feature f.
• T threshold for positive features.
• T- threshold for negative features.
• Second, rank the remaining features based on
  expected entropy loss.

8
Expected Entropy Loss

• The prior entropy of the class distribution
• The posterior entropy of the class when the
  feature is present
• The posterior entropy of the class when the
  feature is absent
• The expected entropy loss

9
Train SVM

• A set of data points (x1,y1),, (xN,yN)
• xi is an input and yi is a target output (1 or
  -1).
• Separating hyperplane
• wf(xi) b 0
• wf(xi) b 1 if yi 1
• wf(xi) b -1 if yi -1
• wf(xi) b
• where
• minimizing
• Output
• Kernel function

10
Improvement-Uncertainty Sampling

• The result from an SVM classifier is a real
  number from -8 to 8.
• When the output is on the interval (-1,1) it is
  less certain than if it is on the intervals
  (-8,-1) and (1,8).
• The region (-1,1) is called
• the uncertain region.
• Uncertainty sampling
• A human judges the
• documents in the
• uncertain region

11
Improvement-Combination

• To combine results from the extended anchortext
  based classifier with the less accurate full-text
  classifier.

12
Name the Cluster

• Using the top ranked features extracted from the
  extended anchotexts virtual documents to name a
  cluster.
• Beliefs
• The words near the anchortexts are descriptions
  of the target documents.
• The top ranked features by expected entropy loss
  are those which occur in many positive
  examples,and few negative ones.

13
Results-classifying

• Anchortext alone is comparable for classification
  purpose with the full-text.
• Classification accuracy is significantly improved
  when using the extended anchortext instead of
  the document full-text.
• Combination method is highly effective for
  improving positive-class accuracy, but reduces
  negative class accuracy.
• Uncertainty sampling required examining only 8
  of the documents on average, while providing an
  average positive class accuracy improvement of
  almost 10 percentage points.

14
Result--Clustering

• The full-text appears comparable to the extended
  anchortext.
• The anchortext alone appears to do a poor job of
  describing the category.

15
Future Work

• To include other features on the inbound web
  pages besides extended anchortext
• To examine the effects of the number of inbound
  links.
• To examine the nature of the category by
  expanding this to thousands of categories.
• To study the effects of the positive set size.