A study of the scaling up capabilities of stratified prototype generation

Prototype generation is an appropriate data reduction process for improving the efficiency and the efficacy of the nearest neighbor rule. Specifically, evolutionary prototype generation techniques have been highlighted as the best performing methods. However, these methods can sometimes be inefficient when the data scale up. In other data reduction techniques, such as prototype selection, an stratification procedure has been successfully developed to deal with large data sets. In this study, we test the combination of stratification with prototype generation techniques, considering data sets with more than 10000 instances. We compare some of the most representative prototype reduction methods and perform a study of the effects of stratification in their behavior. The results, contrasted with nonparametric statistical tests, show that several prototype generation techniques present a better performance than previously analyzed methods.