Experiments on Feature Selection and Document Classification

1
Experiments on Feature Selection and Document
Classification

• Hsin-Chen Chiao
• 7/11/2002

2
Characteristics of KNN

• Flat classification
• Determined by Major
• Ex Blue and Green
• How Many and how Much
• Determined by K

3
KNN Classification Puzzles


4
What We Want

• Hierarchical classification
• Do not determine by major
• K value will suit each case

5
Hierarchical Classification for KNN

• Assumption
• Classification trees are well built.
• Document similarities are well calculated.

6
Basic Calculation in HC for KNNBottom-Up
Calculation
A
A1
A2
A3
A4
A5
Sim(A) 1/n(sim(A1)sim(A2)sim(A5))

• Unbalanced classification tree should
• introduce Other class.

7
How to ClassifyTop-Down Classification

• Using Level 3 to determine Level 2
• Recursive to Level 4 to
• determine Level 3
• Classify and take

Level 1
Level 2
Level 3
8
Determine the Value of K

• Using double the least son number of the Level 2

100
20
300
In this case, K 202 40
9
Experiment Result

• Using STIC(87) as corpus,
• 4118 training data, 1364 testing data
• Level 1 with four classes, Level 2 with forty
• Using flat Knn top one precision about 58
• Using one level hierarchical classification top
  one precision about 61

10
Crucial Point --- Feature Selection

• In all experiments, the weight of feature uses
  Document Frequency.
• DF will suffer some problems in Hierarchical
  Classification.
• High level with low freq feature
• Low level with high freq feature
• Because of the lack of training data, Bi-gram
  will result in Cns superior than Eng

11
Crucial Point --- Feature Selection (II)

• When training data is not enough, Chinese term
  should be selected.
• Article similarity will be abnormally high if
  features in an article are few.
• Removing stopword is not easy.

12
Weight of feature

• DF
• TFiDF
• TFCFiDF (also determined by major)

13
Statistical Report

• Level 2 29.7 classes with less than 20 articles
• Level 3 77.8 classes with less than 20 articles

14
Future Work

• Chinese using terms meaningful
• Different classes with different feature weight
• In flat knn, we could also apply the classify
  and take method.