parallel-feature-selector 0.1.1

Package for parallelized feature selection methods

Parallel Feature Selector

A feature selection module that works in parallel among the processors

Installation

pip install parallel-feature-selector

Dependencies:

• pandas
• scikit-learn
• mpi4py
• arff2pandas
• openpyxl

MPI needs to be installed on the computer. For Windows:

• Install both Microsoft MPI v10.0 and Microsoft MPI SDK
• Set up environment variables -- C:\Program Files (x86)\Microsoft SDKs\MPI -- C:\Program Files\Microsoft MPI\Bin
• Install mpi4py with "conda install -c intel mpi4py" in Anaconda Prompt

Usage (Module)

Example Code:

import parallel_feature_selector as pfs
pfs.bruteForce(data, estimator, 0.2, 42, 10, 10, True, 'outputTimeAnalysis.csv', 'outputSubsetAnalysis.csv')

Function Parameters and Default Arguments:

def bruteForce(data, estimator, testSize=0.2, randomState=42, cv=5, topScoreNo=5, shuffle=True, timeResultFile='outputTimeAnalysis.csv', scoreResultFile='outputSubsetAnalysis.csv')

Parameter	Description
data	pandas.DataFrame with column names and numerical values
esimator	scikit-learn esimator object
testSize	float, test size between 0 and 1 for train_test_split (Default = 0.2)
randomState	int or None, random state for train_test_split (Default = 42)
cv	int, fold number for cross-validation (Default = 5)
topScoreNo	int, specifies how many best-scoring subsets to print (Default = 5)
shuffle	bool, used to shuffle the data set before splitting (Default = True)
timeResultFile	Output file path for elapsed time results in csv format (Default = 'outputTimeAnalysis.csv')
scoreResultFile	Output file path for score and subset results in csv format (Default = 'outputSubsetAnalysis.csv')

Usage (Script)

You can also run the code in terminal with the following arguments:

mpiexec -n [PROCESS_NO] python parallel_feature_selector.py [DATA_SET_PATH] [ESTIMATOR_NAME] [TEST_SIZE] [RANDOM_STATE] [CROSS_VALIDATION] [TOP_SCORE_NO] [SHUFFLE] [TIME_RESULT_PATH] [SCORE_RESULT_PATH]

• [RANDOM_STATE] must be given as integer ('None' can be given in module function)
• [SHUFFLE] must be given as integer, 1 for 'True' and 0 for 'False' ('True' or 'False' in module function)

Example:

mpiexec -n 4 python parallel_feature_selector.py C:\Users\BILGISAYAR\Desktop\AcademicPerformance.csv gnb 0.2 1 10 10 1 outputTimeAnalysis.csv outputSubsetAnalysis.csv

When running the script in terminal:

• .csv .xlsx .arff files can be used for input data set
• Available estimator arguments: gnb, bnb, knnMinkowski, knnEuclidean, svcRbf, svcPoly, logReg, dtEntropy, dtGini, rfEntropy, rfGini

Terminal Argument	Scikit-Learn Estimator
gnb	GaussianNB()
bnb	BernoulliNB()
knnMinkowski	KNeighborsClassifier(n_neighbors=5, metric='minkowski')
knnEuclidean	KNeighborsClassifier(n_neighbors=5, metric='euclidean')
svcRbf	SVC(kernel = 'rbf', gamma='scale')
svcPoly	SVC(kernel = 'poly', gamma='scale')
logReg	LogisticRegression(solver='liblinear', multi_class='auto')
dtEntropy	DecisionTreeClassifier(criterion='entropy', random_state = 1)
dtGini	DecisionTreeClassifier(criterion='gini', random_state = 1)
rfEntropy	RandomForestClassifier(n_estimators=10, criterion='entropy', random_state=1)
rfGini	RandomForestClassifier(n_estimators=10, criterion='gini', random_state=1)