Instance Selection to Improve Gamma Classifier

Abstract

Pre-processing the dataset is an important stage in the Knowledge Discovery in Datasets (KDD) process. Filtering noise through instance selection is a necessary task. With this, the risk to use misclassified and non-representative instances to train supervised classifiers is reduced. This study aims at improving the performance of the Gamma associative classifier, by introducing a novel similarity function to guide instance selection. The experimental results, over 15 datasets, include several instance selection methods, and their influence in the performance of Gamma classifier is analyzed. The effectiveness of the proposed similarity function is tested, obtaining good results according to classifier accuracy and instance retention ratio.