A Two-Phase Model for Learning Rules from Incomplete Data

A two-phase learning strategy for rule induction from incomplete data is proposed, and a new form of rules is introduced so that a user can easily identify attributes with or without missing values in a rule. Two levels of measurement are assigned to a rule. An algorithm for two-phase rule induction is presented. Instead of filling in missing attribute values before or during the process of rule induction, we divide rule induction into two phases. In the first phase, rules and partial rules are induced based on non-missing values. In the second phase, partial rules are modified and refined by the imputation of some missing values. Such rules truthfully reflect the knowledge embedded in the incomplete data. The study not only presents a new view of rule induction from incomplete data, but also provides a practical solution. Experiments validate the effectiveness of the proposed method.