An Extension on ``Statistical Comparisons of Classifiers over Multiple Data Sets'' for all Pairwise Comparisons

摘要

In a recently published paper in JMLR, Demšar (2006) recommendsa set of non-parametric statistical tests and procedures which canbe safely used for comparing the performance of classifiers overmultiple data sets. After studying the paper, we realize that thepaper correctly introduces the basic procedures and some of themost advanced ones when comparing a control method. However, itdoes not deal with some advanced topics in depth. Regarding thesetopics, we focus on more powerful proposals of statisticalprocedures for comparing n × n classifiers. Moreover, weillustrate an easy way of obtaining adjusted and comparablep-values in multiple comparison procedures. color="gray">