Practical Notes on Applying Generalised Stochastic Orderings to the Study of Performance of Classification Algorithms for Low Quality Data

摘要

This paper presents an approach to applying stochastic orderings to evaluate classification algorithms for low quality data. It discusses some known stochastic orderings along with practical notes about their application to classifier evaluation. Finally, a new approach based on fuzzy cost function is presented. The new method allows comparing any two classifiers, but does not require a precise definition of the cost function. All proposed methods were evaluated on real life medical data. The obtained results are very similar to those previously reported but comparatively much weaker assumptions about costs values are adopted.