Signal-to-Noise Ratio feature selection for multi-class classification

Abstract

This research aims to develop the Signal-to-Noise Ratio (SNR) feature selection for multi-class classification. Six benchmark datasets are used to test the performance of the proposed method. The datasets are divided into 2 groups: 1) 3 datasets for 2-class classification problem, and 2) 3 datasets for multi-class classification problem. First of all, the experiment starts with 2-class classification problem. Three feature selection techniques are used to compare the performance: Cosine Correlation (CC), Euclidean Distance (ED) and SNR. Two algorithms: Naïve Bayes and K-Nearest Neighbor (KNN) are used as the classifier with WEKA software. The experimental results show that the SNR offers the best result for 2-class classification. Since the SNR can be only used for the 2-class classification problem, this research tries to develop an SNR based feature selection for multi-class classification. The SNR value is the summation of the SNR which is calculated from comparison results of the group 1 and the rest through to n groups. The experimental results show that the proposed method yields the best performance.