online read us now
Paper details
Number 4 - December 2021
Volume 31 - 2021
A weighted wrapper approach to feature selection
Maciej Kusy, Roman Zajdel
Abstract
This paper considers feature selection as a problem of an aggregation of three state-of-the-art filtration methods: Pearson’s
linear correlation coefficient, the ReliefF algorithm and decision trees. A new wrapper method is proposed which, on the
basis of a fusion of the above approaches and the performance of a classifier, is capable of creating a distinct, ordered
subset of attributes that is optimal based on the criterion of the highest classification accuracy obtainable by a convolutional
neural network. The introduced feature selection uses a weighted ranking criterion. In order to evaluate the effectiveness
of the solution, the idea is compared with sequential feature selection methods that are widely known and used wrapper
approaches. Additionally, to emphasize the need for dimensionality reduction, the results obtained on all attributes are
shown. The verification of the outcomes is presented in the classification tasks of repository data sets that are characterized
by a high dimensionality. The presented conclusions confirm that it is worth seeking new solutions that are able to provide
a better classification result while reducing the number of input features.
Keywords
feature selection, wrapper approach, feature significance, weighted combined ranking, convolutional neural network, classification accuracy