On persistence of empirical risk bias in classification

Abstract

The paper presents a research on empirical risk bias in classification problem. The statistical modeling performed shows that the risk bias dependence on decision class capacity appears to be the same both for the multinomial (discrete) case and for the linear classifier. This result ensures that universal scaling of Vapnik-Chervonenkis bias estimations may be available since such scaling was obtained for a discrete case. To prove, an empirical risk was used as a risk estimator in the comparison of it's volatility (deviation) versus the volatility of leave-one-out estimator is also performed.