MESDIF {SDEFSR}

Multiobjective Evolutionary Subgroup DIscovery Fuzzy rules (MESDIF) Algorithm

Description

Performs a subgroup discovery task executing the MESDIF algorithm.

Usage

MESDIF(
  paramFile = NULL,
  training = NULL,
  test = NULL,
  output = c("optionsFile.txt", "rulesFile.txt", "testQM.txt"),
  seed = 0,
  nLabels = 3,
  nEval = 10000,
  popLength = 100,
  eliteLength = 3,
  crossProb = 0.6,
  mutProb = 0.01,
  RulesRep = "can",
  Obj1 = "CSUP",
  Obj2 = "CCNF",
  Obj3 = "null",
  Obj4 = "null",
  targetVariable = NA,
  targetClass = "null"
)

Arguments

paramFile

The path of the parameters file. NULL If you want to use training and test SDEFSR_Dataset variables

training

A SDEFSR_Dataset class variable with training data.

test

A SDEFSR_Dataset class variable with test data. NULL if you only want to use training data.

output

character vector with the paths where store information file, rules file and quality measures file, respectively.

seed

An integer to set the seed used for generate random numbers.

nLabels

Number of linguistic labels that represents numerical variables.

nEval

An integer for set the maximum number of evaluations in the evolutive process. Large values of this parameter increments the computing time.

popLength

An integer to set the number of individuals in the population.

eliteLength

An integer to set the number of individuals in the elite population.

crossProb

Sets the crossover probability. A number in [0,1].

mutProb

Sets the mutation probability. A number in [0,1].

RulesRep

Representation used in the rules. "can" for canonical rules, "dnf" for DNF rules.

Obj1

Sets the Objective number 1. See Objective values for more information about the possible values.

Obj2

Sets the Objective number 2. See Objective values for more information about the possible values.

Obj3

Sets the Objective number 3. See Objective values for more information about the possible values.

Obj4

Sets the Objective number 4. See Objective values for more information about the possible values.

targetVariable

The name or index position of the target variable (or class). It must be a categorical one.

targetClass

A string specifing the value of the target variable. null for search for all possible values.

Details

This function sets as target variable the last one that appear in SDEFSR_Dataset object. If you want to change the target variable, you can set the targetVariable to change this target variable. The target variable MUST be categorical, if it is not, throws an error. Also, the default behaviour is to find rules for all possible values of the target varaible. targetClass sets a value of the target variable where the algorithm only finds rules about this value.

If you specify in paramFile something distinct to NULL the rest of the parameters are ignored and the algorithm tries to read the file specified. See "Parameters file structure" below if you want to use a parameters file.

Value

The algorithm shows in the console the following results:

1. The parameters used in the algorithm

2. The rules generated.

3. The quality measures for test of every rule and the global results. This globals results shows the number of rules generated and means results for each quality measure.

Also, the algorithms save those results in the files specified in the output parameter of the algorithm or in the outputData parameter in the parameters file.

Additionally a SDEFSR_Rules object is returned with this information.

How does this algorithm work?

This algorithm performs a multi-objective genetic algorithm based on elitism (following the SPEA2 approach). The elite population has a fixed size and it is filled by non-dominated individuals.

An individual is non-dominated when (! all(ObjI1 <= ObjI2) & any(ObjI1 < ObjI2)) where ObjI1 is the objective value for our individual and ObjI2 is the objetive value for another individual. The number of dominated individuals by each one determine, in addition with a niches technique that considers the proximity among values of the objectives a fitness value for the selection.

The number of non-dominated individuals might be greater or less than elite population size and in those cases MESDIF implements a truncation operator and a fill operator respectively. Then, genetic operators are applied.

At the final of the evolutive process it returns the rules stored in elite population. Therefore, the number of rules is fixed with the eliteLength parameter.

Parameters file structure

The paramFile argument points to a file which has the necesary parameters for MESDIF works. This file must have, at least, those parameters (separated by a carriage return):

• algorithm Specify the algorithm to execute. In this case. "MESDIF"

• inputData Specify two paths of KEEL files for training and test. In case of specify only the name of the file, the path will be the working directory.

• seed Sets the seed for the random number generator

• nLabels Sets the number of fuzzy labels to create when reading the files

• nEval Set the maximun number of evaluations of rules for stop the genetic process

• popLength Sets number of individuals of the main population

• eliteLength Sets number of individuals of the elite population. Must be less than popLength

• crossProb Crossover probability of the genetic algorithm. Value in [0,1]

• mutProb Mutation probability of the genetic algorithm. Value in [0,1]

• Obj1 Sets the objective number 1.

• Obj2 Sets the objective number 2.

• Obj3 Sets the objective number 3.

• Obj4 Sets the objective number 4.

• RulesRep Representation of each chromosome of the population. "can" for canonical representation. "dnf" for DNF representation.

• targetVariable The name or index position of the target variable (or class). It must be a categorical one.

• targetClass Value of the target variable to search for subgroups. The target variable is always the last variable. Use null to search for every value of the target variable

An example of parameter file could be:

 algorithm = MESDIF
 inputData = "irisd-10-1tra.dat" "irisd-10-1tst.dat"
 outputData = "irisD-10-1-INFO.txt" "irisD-10-1-Rules.txt" "irisD-10-1-TestMeasures.txt"
 seed = 0
 nLabels = 3
 nEval = 500
 popLength = 100
 eliteLength = 3
 crossProb = 0.6
 mutProb = 0.01
 RulesRep = can
 Obj1 = comp
 Obj2 = unus
 Obj3 = null
 Obj4 = null
 targetClass = Iris-setosa 

@section Objective values: You can use the following quality measures in the ObjX value of the parameter file using this values:

• Unusualness -> unus

• Crisp Support -> csup

• Crisp Confidence -> ccnf

• Fuzzy Support -> fsup

• Fuzzy Confidence -> fcnf

• Coverage -> cove

• Significance -> sign

If you dont want to use a objective value you must specify null

References

• Berlanga, F., Del Jesus, M., Gonzalez, P., Herrera, F., & Mesonero, M. (2006). Multiobjective Evolutionary Induction of Subgroup Discovery Fuzzy Rules: A Case Study in Marketing.

• Zitzler, E., Laumanns, M., & Thiele, L. (2001). SPEA2: Improving the Strength Pareto Evolutionary Algorithm.

Examples

 MESDIF( paramFile = NULL,
        training = habermanTra, 
        test = habermanTst, 
        output = c(NA, NA, NA),
        seed = 0, 
        nLabels = 3,
        nEval = 300, 
        popLength = 100, 
        eliteLength = 3,
        crossProb = 0.6,
        mutProb = 0.01, 
        RulesRep = "can",
        Obj1 = "CSUP", 
        Obj2 = "CCNF",
        Obj3 = "null",
        Obj4 = "null",
        targetClass = "positive"
        )

## Not run: 
Execution for all classes, see 'targetClass' parameter
MESDIF( paramFile = NULL,
        training = habermanTra, 
        test = habermanTst, 
        output = c("optionsFile.txt", "rulesFile.txt", "testQM.txt"),
        seed = 0, 
        nLabels = 3,
        nEval = 300, 
        popLength = 100, 
        eliteLength = 3,
        crossProb = 0.6,
        mutProb = 0.01, 
        RulesRep = "can",
        Obj1 = "CSUP", 
        Obj2 = "CCNF",
        Obj3 = "null",
        Obj4 = "null",
        targetClass = "null"
        )
 
## End(Not run)