Prototype generation based on instance filtering and averaging

Abstract

We propose a new algorithm, called Prototype Generation and Filtering (PGF), which combines the strength of instance-filtering and instance-averaging techniques. PGF is able to generate representative prototypes while eliminating noise and exceptions.We also introduce a distance measure incorporating the class label entropy information for the prototypes. Experiments have been conducted to compare our PGF algorithm with pure instance filtering, pure instance averaging, as well as state-of-the-8irt algorithms such as C4.5 and KNN. The results demonstrate that PGF can significantly reduce the size of the data while maintaining and even improving the classification performance.