Error rate as features added. The cross-validated error rate decreases steadily as features are added according to their ability to increase the accuracy of a naïve Bayesian classifier. Two typical trials are shown. It is possible with a large number of variables to continue to choose ever-larger feature sets with ever-lower error rates, but these larger feature sets are caused by over-fitting and have unstable memberships.
Kuschner et al. BMC Bioinformatics 2010 11:177 doi:10.1186/1471-2105-11-177