WIT 2005 Data Mining

1
WIT 2005Data Mining

• A Genetic Algorithm
• for Text Mining

Guy Desjardins intellagent_at_vif.com Robert
Godin godin.robert_at_uqam.ca Robert
Proulx proulx.robert_at_uqam.ca
2
Agenda

• Text Mining Information Retrieval
• Challenge
• Genetic Model
• Experiment and Results
• Conclusion - Future

3
Agenda

• Text Mining Information Retrieval
• Challenge
• Genetic Model ?
• Experiment and Results
• Conclusion - Future

• Algorithm
• Operators
• Fitness

4
Text Mining Information Retrieval

• Common representation
• Indexing documents ? corpus terms

5
Text Mining Information Retrieval

• Text Mining
• Classification of texts
• Within cluster similarity
• Information Retrieval
• Query Document matching
• Similarity between queries and documents
• Similarity(Di, Dj) cosine (di, dj)

6
Text Mining Information Retrieval

• tf ? idf weighting scheme
• tf term frequencies
• idf inverse document frequencies

7
Challenge

• Text Mining
• Improve the within cluster similarities
• Information Retrieval
• Improve retrieval and stability
• Stability and generalization difficult to reach

8
Challenge

• Approach
• Add co-occurrent terms to text representations
• Problem
• Combinatorial explosion ( 10 26)
• Solution
• Genetic algorithm

9
Genetic Model

• Cycle

10
Genetic Model

• Operators
• Crossover exchanges portions of chromosomes
• Mutation modification of a random gene

11
Genetic Model

• Fitness weights of co-occurrent term sets

12
Experiment

• Collection TREC ZF109
• 22 709 documents
• 19 queries
• 72 983 index terms
• GA parameters
• Pinit 1 000 sets of 2-6 co-occurrent terms
• Hyper-mutation rate 50

• Averages
• 6 documents / index term
• 20 terms / document

13
Results

• Text mining

14
Results

• Information retrieval

15
Conclusion - Future

• Genetic algorithm focused on major sub-domains
• Would improve one cluster in text mining
• Not enough to improve retrieval on large
  collections
• Future work
• Apply Genetic algorithm recursively
• Explore other Fitness functions
• Use the correlated terms for query enhancements