engleski

Explicit Feature Construction and Manipulation for Covering Rule Learning

In rule learning systems features are the main rule building blocks. They are either simple tests of attribute values or complex logical relations representing available domain knowledge. In contrast to practice of many existing classification rule learning systems to construct appropriate features during the rule construction process, we argue that separation of the feature construction and rule construction processes has theoretical and practical justification specifically for covering approaches in two class supervised rule learning. Explicit usage of features enables an 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 set enables construction of complete and consistent rules whenever it is possible, and that the set does not include obviously irrelevant features. Additionally, the concept of feature relevancy is important for effectiveness of rule learning. It this work we illustrate the concept in the coverage space and prove that the relative relevancy has the quality-preserving property in respect to the resulting rules. At the end we show that the transformation from the attribute to the feature space enables novel, theoretically justified handling of unknown attribute values. The same approach enables that estimated imprecision of continuous attributes can be taken into account, resulting in construction of robust features in respect to this imprecision.

machine learning ; rule learning ; features

nije evidentirano