Receiver Operating Characteristics (ROC) graph is a popular way of assessing the performance of classification rules. However, as such graphs are based on class conditional probabilities, they are inappropriate to evaluate the quality of association rules. This follows from the fact that there is no class in association rule mining, and the consequent part of two different association rules might not have any correlation at all. This chapter presents an extension of ROC graphs, named QROC (for Quality ROC), which can be used in association rule context. Furthermore, QROC can be used to help analysts to evaluate the relative interestingness among different association rules in different cost scenarios.
Rule learning has been mainly addressed from two different perspectives: predictive and descriptive tasks. Rule learning in predictive tasks is mainly concerned in generating classification rules that form a classifier. On the other hand, rule generation in descriptive tasks focus in finding all rules over a certain confidence that summarizes the data. However, in a broader sense, rules (either predictive or descriptive) can be considered as an association of two binary variables: the rule antecedent and the rule consequent. Rules of the formantecedent → consequent,where both antecedent and consequent are conjunctions of features (for classification rules, the consequent always refers to a single feature), whereas they do not have features in common. The antecedent is also called left-hand side, premise, condition, tail or body and the consequent is called right-hand side, conclusion or head. Throughout this chapter, we will also use the general notation X → Y (X corresponding to the antecedent and Y to the consequent, respectively) to denote a rule.