Query Optimization by Genetic Algorithms

1
Query Optimization by Genetic Algorithms

• Suhail Owais,
• Pavel Kromer,
• Vaclav Snašel
• Department of Computer Science, VŠB-Technical
  University of Ostrava,
• 17. listopadu 15, Ostrava - Poruba, Czech Republic

2
Outline

1. Introduction
2. Information Retrieval (IR)
3. Genetic Algorithms (GA)
4. Optimization
5. State of art
6. IR and GA
7. Experiments
8. Conclusion
9. Future Work

3
Internet
4
Information Retrieval

• In principle, Suppose there are set of documents
  and a person (user of these documents), the user
  formulates a question (request or query) to which
  the answer is a subset of documents satisfying
  the information need expressed by his question
  Relevant documents.
• Searching for information in documents, for
  document in collection of documents, for metadata
  in documents,
• Searching will be in databases, or in hypertext
  networked databases Internet or intranet.

5
Information Retrieval System - IRS

• IRS concerned
• with responding to the requests of users queries
  for the information seeking text.
• with retrieve all relevant documents to user
  query from a collection of documents, with
  retrieving some of non-relevant as less as
  possible.

6
Retrieved - Relevant Documents to the user Query
Collection of Documents

Relevant Doc.
Relevant Retrieved Doc.
Retrieved Doc.
7
IR Evaluation

• The most Measuring performance of retrieval
  effectiveness are
• Precision the percentage of the retrieved
  documents that are relevant to the user query
• Recall the percentage of the relevant documents
  that are retrieved

8
Genetic Algorithm

• GA used Darwinian Evolution to extract
  optimization strategies nature uses successfully
  and transform them for application in
  mathematical optimization theory to find the
  global optimum in defined phase space
• GA are used in IR problems specially in
  optimizing of a Boolean query.
• GA operators Selection, Fitness function,
  Crossover, and Mutation.

9
GA Flowchart Diagram
Condition Satisfied
Yes
Optimized Query
Encoding
Initialize Population
Evaluate Fitness's
No
Regenerate New Offsprings
End
Selection
Start
Crossover
Mutation
?
10
Optimization

• The procedure or procedures are used to make a
  system or design as effective or functional as
  possible, especially the mathematical techniques
  involved.
• Is the process of modifying a system to improve
  its efficiency. The system can be a single
  computer program, a collection of computers  or
  even an entire network such as the Internet.

11
State of the art 1
Evolutionary Learning of Boolean Queries by
Multiobjective Genetic Programming

• Authors Cordon et al., Springer-Verlag GmbH 2002
• Subject Automatic derivation of Boolean queries,
  by incorporating a Pareto-based multiobjective
  evolutionary approach, MOGA, into genetic
  programming technique.
• Notes
• A query represented as a parse tree with maximum
  of 20 nodes.
• Boolean operators used are AND, OR and NOT.
• Maximum number of documents is 1400.
• Result The proposed approach has performed
  appropriately in seven queries of the well known
  Cranfield collection in terms of absolute
  retrieval performance and of the quality of the
  obtained Paretos.

12
State of the art 2
An Appropriate Boolean Query Reformulation
Interface for Information Retrieval Based on
Adaptive Generalization

• Authors Yoshioka et al., WIRI 2005, In
  Conjunction with IEEE 2005, Tokyo Japan
• Subject Implement a user query interface that
  supports reformulation of IR queries by using
  abstract concepts.
• Notes
• IR interface uses small numbers of query terms
  and concept categories with Boolean expression.
• Reformulate a Boolean query by using only words
  that exist in the original query.
• Boolean operators used are AND, and OR.
• Result Proposed a new IR interface with Boolean
  query reformulation (ABRIR-AG). Find
  complementary query terms that exist in relevant
  documents and reformulate Boolean query formulas
  to clarify the information need.

ABRIR-AG Appropriate Boolean query
Reformulation for IR- Adaptive Generalization
13
IR and GA

• Collection or set of Documents
• Terms for Document di
• Weighting function

T e r m s T e r m s T e r m s T e r m s T e r m s T e r m s T e r m s T e r m s T e r m s
W1 W2 W3 W4 W5 W6 W7 W8 W9
D O C U M E N T s d1 1 1 0 1 0 1 1 0 0
D O C U M E N T s d2 1 1 0 1 1 1 1
D O C U M E N T s d3 0 0 1 1 0 1 1 0 1
D O C U M E N T s d4 1 0 1 1 1 0 1 0 1
1
0
W2 Not in Document d2
W8 IN Document d2
14
Chromosome Encoding

• A query combination from set of terms and set of
  Boolean operators
• Set of queries will be encoded to be chromosomes
  for genetic programming in prefix form such as

(w2 OR w6) AND (w9 AND w3)
AND (OR w2 w6) (AND w9 w3)
? Prefix ?
(w3 AND w4) XOR ((w5 AND w6) OR w8)
XOR (AND w3 w4) (OR (AND w5 w6) w8)
? Prefix ?
15
Tree Structure Representation
XOR (AND w3 w4) (OR (AND w5 w6) w8)
AND (OR w2 w6) (AND w9 w3)
16
Fitness function

• Recall and Precision functions are used to
  Evaluate Chromosomes.

• Selection Operators
• From the population of chromosomes, the best two
  chromosomes depending on the highest fitness
  values for precision or recall measures will be
  selected.

• rd the relevance of document d (1 for relevant
  and 0 for nonrelevant),
• fd the retrieved document d (1 for retrieval
  and 0 for nonretrieval), and
• a and ß are arbitrary weights added specially to
  precision fitness function.

17
Crossover Operator

• Chose Randomly one node position in each Tree to
  be exchanged

OR 4
OR 1
18
Exchange Sub trees
Created two New Offsprings
19
ReIndexing nodes in Offsprings
20
Mutation Operator
Randomly will change one of the Boolean logical
operators to another and the position randomly
chosen
AND , OR , XOR
AND , XOR
AND
No one select, SO no mutation will be done over
this offspring
AND 4
21
Experiments

• Implementation for our Genetic Program was
  tested under the following conditions and
  limitations-
• Two sets of queries that represent in a tree
  prefix forms used as two different initial
  populations
• Boolean model of a collection of documents
• Different Collections of documents
• User query / request
• w8 OR w2

22
Initial Populations

• The two initial population differs by containing
  sub queries, so
• Initial Population 1 contain sub query
• w8 AND w2
• Initial Population 2 contains sub queries
• w8 AND w2
• w8 OR w2
• w8 XOR w2

Initial Population 2
Initial Population 1
1. ("w13"and"w8")and("w10"or"w4")
2. ("w1"and("w8"and"w2"))or("w4"or"w2")
3. ("w1"or"w2")and(("w5"or"w4")and("w3"and"w6"))
4. ("w9"and"w14")
5. ("w14") and "w1"
6. ("w2"or"w6")or("w8"and"w13")
7. ("w3"And"w4")or(("w12" xor "w15")and"w8")
8. "w1" or "w5"
1. ("w13"and"w8")and("w10"or"w4")
2. ("w1"and("w8"and"w2"))or("w4"or"w2")
3. ("w1"or"w2")and(("w5"or"w4")and("w3"and"w6"))
4. ("w9"and"w14")
5. ("w14") and "w1"
6. ("w2"xor"w8")or("w8"and"w13")
7. ("w3and"w4")or(("w2" or "w8")and"w8")
8. "w1" or "w5"
23
Variables initialization

• Crossover probability value ? 0.8
• Mutation probability value ? 0.2
• Population size (number of chromosomes) ? 8
• Maximum number of generations ? 50.
• a ? 0.25
• ß ? 1.0

24
Document Collections

• Three different document collections with variant
  number of words and documents.

Collection Number of Documents Number of Words Maximum Number of Words in each Document
1 10 30 10
2 200 50 50
3 5000 2000 500
25
Notes on limitations

• Single point for Crossover
• Mutation operator applied only over Boolean
  operators AND, OR or XOR.
• Fitness operator must be defined in input data to
  be-
• PrecisionFitness or
• RecallFitness.
• maximum value for PrecisionFitness a ß
• so It may be grater than one ( gt 1 )
• it can not be interpreted as the probability of
  retrieving relevant document.

26
Experiments

• Set of experiments done over three test cases
  depends on.
• Initial Population used
• Initial Population 1 OR
• Initial Population 2.
• Fitness function used
• PrecisionFitness OR
• RecallFitness.
• Collection used
• Collection 1 OR
• Collection 2 OR
• Collection 3.

27
Experiments Results Using IP. 1
IP Initial Population, FF Fitness Function , R
Recall , P Precision, V Value
Collection 1 Collection 1 Collection 1 Collection 1
FF Relaxed Optimized Query P. V. R.V.
P. (( "w8" OR "w2" ) OR ( "w8" AND "w2" )) 1.25 1.00
R. (( "w13" OR "w8" ) OR ( "w6" OR "w2" )) 1.08 1.00
Collection 2 Collection 2 Collection 2 Collection 2
P. (( "w13" AND "w8" ) OR ( "w8" OR "w2" )) 1.25 1.00
R. ( "w8" OR "w2" ) 1.25 1.00
Collection 3 Collection 3 Collection 3 Collection 3
P. ( "w13" AND "w8" ) 1.05 0.18
R. ( "w8" OR "w2" ) 1.25 1.00
28
Experiments Results Using IP. 2
IP Initial Population, FF Fitness Function , R
Recall , P Precision, V Value
Collection 1 Collection 1 Collection 1 Collection 1
FF Relaxed Optimized Query P. V. R.V.
P. (("w8 OR "w2")OR(("w2 AND "w4")OR(("w8 OR "w2")AND "w1"))) 1.25 1.00
R. (( "w2" OR "w4" ) OR ( "w6" OR "w2" )) 1.07 0.93
Collection 2 Collection 2 Collection 2 Collection 2
P. (( "w13" AND "w8" ) OR ( "w8" XOR "w2")) 1.25 1.00
R. (( "w13" AND "w8" ) OR ( "w8" XOR "w2" )) 1.25 1.00
Collection 3 Collection 3 Collection 3 Collection 3
P. ( "w8" OR "w2" ) 1.25 1.00
R. (( "w13" AND "w8" ) OR ( "w8" OR "w2" )) 1.25 1.00
29
Precision and Recall Diagrams
Collections
30
Precision and Recall Diagrams
Initial Populations
31
Conclusions

• The final population contains set of individuals
  that have same fitness values
• one randomly chosen to be an optimized query.
• Because of selection queries with different sub
  queries similar to the user query that increase
  the quality of the initial population selected
• ? this obtained better results
• Especially when precision was used as fitness
  measure and experiment was done over largest
  collection, the fitness values of recall in final
  population were low.
• in many experiments mostly all members of
  population reached the maximum values of
  precision and recall before reaching given number
  of generations.

32
Future works

• Use more of unweighted Boolean operators like
• ( ADJ, and OF) operators
• Mutation operates over all Boolean operators
  (AND, OR, XOR, ADJ, OF, and NOT)
• Try to improve selection method for choosing the
  best individual from a set of queries with equal
  values of precision or recall.
• Appling of fuzzy theorem approach in this
  problematic
• - Use weights for terms in documents instead of
  Boolean weights.

33
Thanks for your attention
Suhail Owais suhailowais_at_yahoo.com