A novel Supervised Instance Selection algorithm

Abstract

Instance selection is often used in case of lazy classifiers. This paper addresses the need of instance selection in case of neural network and decision tree classifiers and presents a novel Supervised Instance Selection (SIS) algorithm. Initially, a neural network classifier is constructed using all training instances. The algorithm then selects a few instances using the certainty values of the wrapped neural network to construct a compact classifier. Empirical study made with standard datasets shows that SIS save on 70% of storage space without degrading the accuracy. It is independent of nature of the dataset and the tool used.