Abstract
In the association rule mining field many different quality measures have been proposed over time with the aim of quantifying the interestingness of each discovered rule. In evolutionary computation, many of these metrics have been used as functions to be optimized, but the selection of a set of suitable quality measures for each specific problem is not a trivial task. The aim of this paper is to review the most widely used quality measures, analyze their properties from an empirical standpoint and, as a result, ease the process of selecting a subset of them for tackling the task of mining association rules through evolutionary computation. The experimental analysis includes twenty metrics, thirty datasets and a diverse set of algorithms to describe which quality measures are related (or unrelated) so they should (or should not) be used at time. A series of recomendations are therefore provided according to which quality measures are easily optimized, what set of measures should be used to optimize the whole set of metrics, or which measures are hardly optimized by any other.