A new bootstrapping method to improve classification performance in learning classifier systems

Abstract

A new technique for improving the classification performance of learning classifier systems (LCS) was developed and applied to a real-world data mining problem. EpiCS, a stimulus-response LCS, was adapted to perform prevalence-based bootstrapping, wherein data from training and testing sets were sampled according to the prevalence of the individual classes, rather than randomly using the class distribution inherent in the data. Prevalence-based bootstrapping was shown to improve classification performance significantly on training and testing. Furthermore, this procedure was shown to enhance EpiCS’s classification performance on testing compared to a well-known decision tree inducer (C4.5) when similar bootstrapping procedures were applied to the latter.