Extending Rule-based Classifiers to Improve Recognition of Imbalanced Classes

This papers deals with inducing rule-based classifiers from imbalanced data, where one class (a minority class) is under-represented in comparison to the remaining classes (majority classes). We discuss reasons for bias of standard classifiers toward recognition of examples from majority classes and misclassifcation of the minority class. To avoid limitations of sequential covering approaches, we present a new approach to improve sensitivity of the rule based classifier. It includes a modification the structure of sets of rules, where for majority classes minimal sets of rules are still induced while the rule set for the minority class is generated by the algorithm, called EXPLORE. This algorithm produces rules being more general and supported by more learning examples than rules from the minimal set. The usefulness of the new approach is verified in a comparative experiments on several imbalanced data sets.