Explicit Feature Construction and Manipulation for Covering Rule Learning Algorithms

Abstract

Features are the main rule building blocks for rule learning algorithms. They can be simple tests for attribute values or complex logical terms representing available domain knowledge. In contrast to common practice in classification rule learning, we argue that the separation of feature construction and rule construction processes has a theoretical and practical justification. Explicit usage of features enables a unifying framework of both propositional and relational rule learning and we present and analyze procedures for feature construction in both types of domains. It is demonstrated that the presented procedure for constructing a set of simple features has the property that the resulting feature set enables the construction of complete and consistent rules whenever possible, and that the set does not include obviously irrelevant features. It is also shown that feature relevancy may improve the effectiveness of rule learning. It this work, we illustrate the relevancy concept in the coverage space, and show that the transformation from the attribute to the feature space enables a novel, theoretically justified way of handling unknown attribute values. The same approach enables that the estimated imprecision of continuous attributes can be taken into account, resulting in the construction of features that are robust to attribute imprecision.