Improving Web Clustering by Cluster Selection

1
Improving Web Clustering by Cluster Selection
Daniel Crabtree, Xiaoying Gao, Peter
Andreae Victoria University of Wellington, New
Zealand
2
Web Search
2

• Iterative Process
• Problems with Standard Web Search
• Many Irrelevant Results
• Single Long List
• Solution
• Identify and Present Implicit Clusters

3
Web Clustering
3
Search Results for Jaguar 1 6 of
70,000,000
1. Jaguar Official worldwide web site of Jaguar
Cars. 2. Apple - Mac OS X The Apple Mac OS X
product page. 3. Jaguar UK - R is for Racing
The essence of the Jaguar breed 4.
Jaguar General information from Big Cats Online.
5. Jaguar AU - Jaguar Cars Services and
news 6. Jaguar -- Defenders of Wildlife Size,
appearance, life span and diet.
4. Jaguar General information from Big Cats
Online. 6. Jaguar -- Defenders of
Wildlife Size, appearance, life span and diet.
Clusters 1. Car 2. Animal 3. Mac OS 4. Other
4
Web Clustering Algorithms
4

• Many standard clustering algorithms.
• Text oriented clustering algorithms
• STC - Suffix Tree Clustering
• ESTC - Improvement on STC

5
Suffix Tree Clustering
5
Reference Zamir and Etzioni
6
STC Identify Base Clusters
6
7
STC Combining Base Clusters
7
Merge Clusters Based On Overlap
30
18
7
12
6
Merged Cluster Score is sum of base cluster scores
8
STC Rank/Select Clusters
8

• Sort Clusters by Score
• Select Best N

9
Problems with STC
9

• STC is better than many other algorithms
• BUT not good enough
• Scores
• Poor Cluster Quality Measure
• Selection
• Poor Coverage
• Excessive Overlap

10
ESTC Better Cluster Scoring
10

• Base Cluster Scores OK
• Combined Cluster Scores BAD
• Overlap between clusters over counted in sum
• Example - Particularly Similar Pages

11
ESTC Scoring Solution
11

• Solution
• Eliminate the over counting of the overlap
• Merged Cluster Score
• Sum over document scores
• Document Score
• Average phrase score of base clusters containing
  the document in the merged cluster

12
ESTC Better Cluster Selection
12

• Top N Clusters BAD
• Dominant Topic over represented

13
ESTC Smarter Selection The Search
13

• ESTC Smarter selection
• Heuristic
• Minimize Overlap
• Maximize Coverage

14
ESTC The Search
14

• Incremental
• Greedy
• Look-ahead Protection
• Sophisticated Branch and Bound Pruning

15
Evaluation Method
15

• Gold Standard - Ideal Clustering
• 2 Searches and 2 Types of Input Data
• Jaguar and Salsa
• Snippets and Full Text
• Precision
• Cluster accuracy against the best matching ideal
  cluster
• Recall
• Coverage of ideal cluster in matched clusters
• F-measure
• Combination of precision and recall

16
Results STC, STC-NS, ESTC
16
Jaguar Full Text Clustering Results
17
Results ESTC vs Grokker
17

• Similar performance without page titles
• Page titles are often very useful
• Algorithm Input F-measure
• ESTC Snippets 58
• Grokker Snippets Page Titles 62
• ESTC Full Text 74

18
Conclusions
18

• ESTC has
• A new cluster scoring
• A new cluster selection algorithm
• ESTC is better than STC, and compares favourably
  with Grokker.
• ESTC Scoring function applicable to any
  agglomerative clustering algorithm.
• ESTC Cluster Selection algorithm more widely
  applicable.

19
Future Work
19

• Make improvements to other stages of STC
• Particularly Combining Base Clusters
• Apply cluster selection method to other
  algorithms
• Improve cluster selection heuristic