1. Identifying Interesting Association Rules with Genetic Algorithms
Elnaz Delpisheh
York University
Department of Computer Science and Engineering
April-10-17

2. Data mining I = {i1,i2,...,in} is a set of items.
Too much data
Data
Data Mining
I = {i1,i2,...,in} is a set of items.
D = {t1,t2,...,tn} is a transactional database.
ti is a nonempty subset of I.
An association rule is of the form AB, where A and B are the itemsets, A⊂ I, B⊂ I, and A∩B=∅ .
Apriori algorithm is mostly used for association rule mining.
{milk, eggs}{bread}.
Association rules
There exist other algorithms apart from apriori such as Fp-growth.

3. Apriori Algorithm TID List of item IDs T100 I1,I2,I3 T200 I2, I4 T300
I1, I2, I3, I5
T900
I1, I2, I3

4. Apriori Algorithm (Cont.)

5. Association rule mining
Too much data
Data
Data Mining
Too many association rules
Association rules

6. Interestingness criteria
Comprehensibility.
Conciseness.
Diversity.
Generality.
Novelty.
Utility.
...

7. Interestingness measures
Subjective measures
Data and the user’s prior knowledge are considered.
Comprehensibility, novelty, surprisingness, utility.
Objective measures
The structure of an association rule is considered.
Conciseness, diversity, generality, peculiarity.
Example: Support
It represents the generality of a rule.
It counts the number of transactions containing both A and B.

8. Drawbacks of objective measures
Detabase-dependence
Lack of knowledge about the database
Threshold dependence
Solution
Multiple database reanalysis
Problem
Large number of disk I/O
Problem with Multiple database reanalysis is that, some databases are simply large. Association rule mining must confront exponential search spaces.
Detabase-independence
This approach does not require users to specify thresholds. Instead of generating unknown number of interesting rules like the traditional models, only the most interesting rules are extracted according to the interestingness measure as defined by the fitness function!
Detabase-independence

9. Genetic algorithm-based learning (ARMGA )
Initialize population
Evaluate individuals in population
Repeat until a stopping criteria is met
Select individuals from the current population
Recombine them to obtain more individuals
Evaluate new individuals
Replace some or all the individuals of the current population by off-springs
Return the best individual seen so far
Usually genetic algorithm for rule mining are divided into 2 groups according to their encoding of rules in the population of chromosomes.
-Michigan Approach
Many ppl have used this approach. However, if the number of rules is too many, this approach is impractical.
-Pittsburgh Approach

10. ARMGA Modeling Given an association rule XY Requirement
Conf(XY) > Supp(Y)
Aim is to maximise
Conf(XY) > Supp(Y), since we are only interested in positive rules.

11. ARMGA Encoding Michigan Strategy
Given an association k-rule XY, where X,Y⊂I, I is a set of items I=i1,i2,..., in, and X∩Y=∅.
For example
{A1,...,Aj}{Aj+1,...,Ak}
Michigan Approach
Each rule is encoded into an individual
Pittsburgh Approach
A set of rules are encoded into a chromosome.

12. ARMGA Encoding (Cont.)
The aforementioned encoding highly depends on the length of the chromosome.
We use another type of encoding:
Given a set of items {A,B,C,D,E,F}
Association rule ACFB is encoded as follows
00A11B00C01D11E00F
00: Item is antecedent
11: Item is consequence
01/10: Item is absent

13. ARMGA Operators
Select
Crossover
Mutation

14. ARMGA Operators-Select
Select(c,ps): Acts as a filter of the chromosome
C: Chromosome
Ps: pre-specified probability

15. ARMGA Operators-Crossover
This operation uses a two-point strategy

16. ARMGA Operators-Mutate

17. ARMGA Initialization

18. ARMGA Algorithm

19. Empirical studies and Evaluation
Implement the entire procedure using Visual C++
Use WEKA to produce interesting association rules
Compare the results