数値属性からの例外ルール発見

抄録

Extracting classification rules from pre-classified instances is an important field of knowledge discovery. Typically, enormous number of classification rules are embedded in the given instances and a discovery system has to extract only useful rules effectively. There are many types of useful rules depending on what the discovered rules are used for and this paper deals with discovery of an exception rule in which each condition in its bodies has no or negative correlation with its conclusion (class label) but the combination of the conditions predicts the conclusion accurately. Exception rules are important because such hidden relationships may be unknown, new knowledge for human experts. In this paper, we first propose a new criterion that judges whether a body of a rule to be evaluated is realy required to predict its conclusion by comparing its accuracy with those of more general rules. We also propose a discovery algorithm, IIS, to extract exception rules with respect to the above criterion from training instances represented by numeric attributes. IIS effectively searches exception rules by iteratively specializing possible intervals of attributes in a rule's body. The empirical results with UCI repository show that IIS is sufficiently fast to extract exception rules with three or less attributes in their bodies.