Learning Composite Concepts

Abstract

Some recent work investigates concept learning in the context of description logics. Given the fact that first-order logic has been restricted in several ways for its use in the field of machine learning, description logics seem to make another good candidate as a learning framework. However, it is only recently that learning algorithms have been developed within this framework [CH94a, CH94b, FP94, KM94] and we are not aware of any learning algorithms (learning in description logics or other kinds of learning) that incorporate special knowledge about the part-of relation. In this chapter we sketch how we can use our description logic for composite objects to include extra information about this important relation into learning. On one hand we have a relatively rich representation language with an infinite space of possible concepts. On the other hand we have special constructs for handling part-of relations that can be used in the learning algorithm to reduce the overall search space. In section 10.2 we describe the framework that we use and we define the learning task in section 10.3. Two useful operations in learning in description logics are to compute a least common subsumer for two concepts and to associate a concept with an individual that captures as much information about that individual as possible. We describe these operations for our framework in section 10.4. The algorithms are described informally in section 10.5.