A Novel Approach to Compute Confusion Matrix for Classification of n-Class Attributes with Feature Selection

Abstract

Confusion matrix is a useful tool to measure the performance of classifiers in their ability to classify multi-classed objects. Computation of classification accuracy for 2-classed attributes using confusion matrix is rather straightforward whereas it is quite cumbersome in case of multi-class attributes. In this work, we propose a novel approach to transform an n × n confusion matrix for n-class attributes to its equivalent 2 × 2 weighted average confusion matrix (WACM). The suitability of WACM has been shown for a classification problem using a web service data set. We have computed the accuracy of four classifiers, namely, Naïve Bayes(NB), Genetic Programming(GP), Instance Based Lazy Learner(IB1), and Decision Tree(J48) with and without feature selection. Next, WACM has been employed on the confusion matrix obtained after feature selection which further improves the classification accuracy.