Abstract
Filters represent a class of feature selection methods used to select a subset of useful features from high
dimensional data on the basis of relevance and redundancy analysis. Maximum relevance minimum redundancy
(mRMR) is a famous feature selection algorithm for microarray data [1]. The quotient based version of
maximum relevance minimum redundancy (Q-mRMR) filter [1],[2] selects, at each iteration, the feature scoring
maximum ratio between its class relevance and average redundancy over already selected subset. This ratio can
be surprisingly large if the denominator i.e. redundancy term is very small, hence suppressing the effect of
relevance and leads to the selection of features which can be very weak representatives of the class. This paper
addresses this issue by presenting a maximum relevance maximum antiredundancy (mRmA) filter method. For
mRmA the value of objective function is within reasonable limits for all values of relevance and redundancy,
hence, making selection of appropriate features more probable. Our 10 fold cross validation accuracy results
using naive Bayes and support vector machines (SVM) classifiers confirm that the proposed method
outperforms both Q-mRMR and Fast Correlation based Filter (FCBF) methods on six datasets from various
applications like microarray, image and physical domains.
Abdul Mannan, Kashif Javed, Serosh Karim Noon. (2017) Maximum Relevance Maximum Anti-Redundancy (mRmA) Feature Selection, Pakistan Journal of Engineering and Applied Sciences, VOLUME 21, Issue 1.
-
Views
2125 -
Downloads
173
Next Article
Article Details
Volume
Issue
Type
Language
