Title: A new feature extraction technique for classifiers using self-organising map

Authors: Prasenjit Dey; Tandra Pal

Addresses: Department of Computer Science and Engineering, National Institute of Technology Durgapur, Durgapur 713209, India ' Department of Computer Science and Engineering, National Institute of Technology Durgapur, Durgapur 713209, India

Abstract: Neural network classifiers often suffer from the overfitting problem which reduces its generalisation capability. The objective of the proposed work is to improve the generalisation of the classifiers by improving the input space of the dataset by self-organising map (SOM) based feature extraction technique. After the training of the SOM network, a Gaussian function is used over the Euclidean distance between the input pattern and the weight vector corresponding to each node in SOM output map. It produces m² dimensional new representation corresponding to each input pattern, where m² is the number of nodes present in the output map. Thereafter, classifiers like probabilistic neural network (PNN) or multilayer perceptron (MLP) is used over this new representation of the input patterns. We have used 12 standard classification datasets to compare the proposed model with conventional PNN and MLP classifiers. Comparison results show the superiority of the proposed method.

Keywords: feature extraction; Gaussian function; generalisation; multilayer perceptron; MLP; probabilistic neural network; PNN; self-organising map; SOM.

DOI: 10.1504/IJCONVC.2016.090079

International Journal of Convergence Computing, 2016 Vol.2 No.3/4, pp.208 - 219

Accepted: 17 Dec 2016
Published online: 28 Feb 2018 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article