Induction of rules subject to a quality constraint: probabilistic inductive learning

• Published in: IEEE transactions on knowledge and data engineering Vol. 5; no. 6; pp. 979 - 984
• Main Authors: Gur-Ali, O., Wallace, W.A.
• Format: Journal Article
• Language: English
• Published: IEEE 01.12.1993
• Subjects: Classification tree analysis; Decision trees; Guidelines; Induction generators; Knowledge acquisition; Machine learning; Machine learning algorithms; Partitioning algorithms; Quality control
• ISSN: 1041-4347
• DOI: 10.1109/69.250081

Organizational databases are being used to develop rules or guidelines for action that are incorporated into decision processes. Tree induction algorithms of two types, total branching and subset elimination, used in the generation of rules, are reviewed with respect to their treatment of the issue of quality. Based on this assessment, a hybrid approach, probabilistic inductive learning (PrIL), is presented. It provides a probabilistic measure of goodness for an individual rule, enabling the user to set maximum misclassification levels, or minimum reliability levels, with predetermined confidence that each and every rule will satisfy this criterion. The user is able to quantify the reliability of the decision process, i.e., the invoking of the rules, which is of crucial importance in automated decision processes. PrIL and its associated algorithm are described. An illustrative example based on the claims process at a workers' compensation board is presented.< >