A Search Engine That Learns

1
A Search Engine That Learns

• Jeff Elser jelser_at_cs.montana.edu
• John Paxton paxton_at_cs.montana.edu
• Montana State University - Bozeman

2
Presentation Outline

• Problem
• Background Information
• Approach
• Preliminary Results
• Future Work
• Summary
• Questions

3
I. Problem

• RightNow software use
• Spidering and searching
• Website optimization
• Page by page is tedious and time consuming
• Dual ownership should allow perfect optimization
• Solutions
• Search engine adjustments
• Suggesting specific web page changes

4
II. Background Search Engine

• Spidering
• Indexing
• Weighting factors

5
II. Background Genetic Algorithms

• Goldbergs Simple GA
• Mutation
• Crossover
• Elitism
• Non-overlapping populations
• Several fitness functions

• Individual 1
• Fitness 2
• Individual 2
• Fitness 4

6
III. Approach

• Architecture
• Training data
• Testing controls (website source)
• GA specifics
• Fitness functions

7
A. Architecture
8
B. Training Data

• Website source
• 20000 newsgroup articles from UCI Knowledge
  Discovery in Databases Archive
• Hand formatted HTML
• Chosen for word count and structure

9
C. Testing Controls

• Webmaster provides training data
• List of important keywords
• Associated ranked pages
• Tedious, but trivial compared to optimizing all
  pages

10
D. GA Specifics

• Random initial population
• Population size 1000
• Used GAlibs built in random number generator
• Genome
• 16 real numbers corresponding to the 16 weighting
  factors
• Range 0.0 1000.0

11
D. GA Specifics

• GA executes for 10000 generations
• Elitism is turned on
• Mutation probability 0.01
• Crossover probability 0.6

12
D. Fitness Function 1

• ?D
• D (actual ranking) (desired ranking)
• 1 to avoid division by 0

13
D. Fitness Function 2

• 100 penalty for pages that dont appear
• -10 reward for pages with a perfect fit

14
IV. Preliminary Results

• 12 tests using fitness function 2
• 1 realistic set of desired rankings
• 11 random sets
• 4 tests obtained perfect rankings
• 4 improved rankings, but did not achieve optimal
• 4 tests showed no improvement

15
IV. Preliminary Results
Htdig default weights
Fitness Function 2
16
IV. Preliminary Results
Fitness Function 2
Htdig default weights
17
V. Future Work Fitness Function 3Levenshtein
Distance

• D string 1 A string 2
• Construct a mxn Matrix (M) where m D1 and n
  A1
• M0,i i and Mj,0 j
• For each remaining cell
• Di Aj then cost 0
• Di ! Aj then cost 1
• Mi,j MIN a, b, c where
• a Mi-1,j 1
• b Mi,j-1 1
• c Mi-1,j-1 cost
• Distance Mm,n

M
R
A
F
4
3
2
1
0
3
2
1
0
1
F
2
1
1
1
2
R
2
2
2
2
3
O
2
3
3
3
4
M
18
V. Future Work Fitness Function 3Levenshtein
Distance

• Reduce the url comparison to string comparison
• Experiment further using LD as a fitness function
• Sigmoid weighting function to increase the
  importance of the front of the string

?
19
V. Future Work

• Create more extensive test sets
• dare.com, studentaid.ed.gov, fafsa.ed.gov,
  americorps.org

20
V. Future Work
21
V. Future Work
22
V. Future Work

• Dare.com
• Use of Meta tags and title tag vary only
  occasionally from page to page
• Keyword Marijuana
• Desired rankings from an index show a main
  background page, and a link to a medical
  marijuana page
• Actual rankings from the local site search put
  the medical marijuana page 3nd, and the
  background page 7th
• Htdig with default weights ranked them both worse
  than 20th

23
V. Future Work

• For pages that still do not rank properly, create
  optimization suggestions
• Use custom meta tags to properly rank outliers
• Use implicit user feedback to find the desired
  rankings

24
VI. Summary

• Proof of concept
• Testing on real world websites will strengthen
  results and open other areas of study.

25
VII. Questions

• Thanks for attending
• Any questions?