The quality of an association pattern can be evaluated using many different measures, such as confidence, support, and interest. This makes it difficult to select the appropriate measure for a particular application. In this paper, the authors analyze the properties and correlations of 21 evaluation measures. First, they exhaustively enumerate whether each one satisfies eight mathematical properties. These properties highlight the main conceptual differences between the measures, and explain why they behave differently in general, and why they become more consistent when they are applied to associations that have a minimum support threshold. Additionally, the authors discuss their behavior after a standardization of the contingency table, and propose a standardization procedure.
The experimental part of the paper has a more methodological goal. The authors develop a way for an expert to choose the measure that is most appropriate for a specific situation. To do this, the expert has to rank a small number of representative contingency tables, and then select the measure that ranks the tables most similarly.
The work is restricted to measures for evaluating nonoriented frequent item sets with only two variables, so it is not extensible to general association patterns, or to oriented association rules. Nonetheless, the work is a very good reference, with a thorough conceptual and experimental analysis of 21 evaluation measures for association patterns. In addition, it proposes a practical methodology for choosing the measure that subjectively could be most adequate in a practical scenario according to an expert.