Micro and macro evaluation of classification rules

Rule evaluation plays an important role in the rule learning and classification process. Many existing rule inductive learning algorithms are based on single rule evaluation measures. However, the overall rule induction system performance and the classification process are involved with the evaluation of a set of rules. This brings the needs for studying the connections between single rule and rule set evaluation measures. The main objective of this paper is to introduce a general framework of classification rule evaluation which connects two types of evaluations, called micro and macro evaluation. Micro evaluation is based on single rules which can be measured by the common empirical measures. Macro evaluation is based on rule sets, depending on the relationships between rules in the set, different resolutions can be applied. By analyzing the relationships between these two types of evaluations, we suggest that under certain conditions, macro evaluation measures can be explicitly expressed by micro evaluation measures.