A Representation for Accuracy-Based Assessment of Classifier System Prediction Performance

Abstract

The increasing use of learning classifier systems (LCS) in data mining necessitates a methodology for improving the assessment of predictive accuracy at both the individual classifier and system levels. A metric, predictive value, is used extensively in clinical diagnosis and medical decision making, and is easily adapted to the LCS to facilitate assessing the ability of individual classifiers used as rules to predict class membership. This metric can also be used to assess the ability of a trained LCS to predict the class of unseen cases. Positive and predictive values were incorporated into an existing LCS model, EpiCS, and applied to 6-Multiplexer data and a sample data set drawn from a large hospitalization survey. The predictive performance of EpiCS on the hospitalization data was shown to be comparable to that of logistic regression and decision tree induction.