Statistical Comparisons of Classifiers by Generalized Stochastic Dominance

Although being a question in the very methodological core of machine learning, there is still no unanimous consensus on how to compare classifiers. Every comparison framework is confronted with (at least) three fundamental challenges: the multiplicity of quality criteria, the multiplicity of data sets and the randomness/arbitrariness of the selection of data sets. In this paper, we add a fresh view to the vivid debate by adopting recent developments in decision theory. Our resulting framework, based on so-called preference systems, ranks classifiers by a generalized concept of stochastic dominance, which powerfully circumvents the cumbersome, and often even self-contradictory, reliance on aggregates. Moreover, we show that generalized stochastic dominance can be operationalized by solving easy-to-handle linear programs and statistically tested by means of an adapted two-sample observation-randomization test. This indeed yields a powerful framework for the statistical comparison of classifiers with respect to multiple quality criteria simultaneously. We illustrate and investigate our framework in a simulation study and with standard benchmark data sets.