Improved Comprehensibility and Reliability of Explanations via Restricted Halfspace Discretization

摘要

A number of two-class classification methods first discretize each attribute of two given training sets and then construct a proposi-tional DNF formula that evaluates to True for one of the two discretized training sets and to False for the other one. The formula is not just a classification tool but constitutes a useful explanation for the differences between the two underlying populations if it can be comprehended by humans and is reliable. This paper shows that comprehensibility as well as reliability of the formulas can sometimes be improved using a discretization scheme where linear combinations of a small number of attributes are discretized.