Probabilistic Concept Formation in Relational Domains

Abstract

This chapter describes COBWEBR’ s representation of objects and concepts along with its organization of concepts in memory. COBWEBR is based on Fisher's (1987) C O B W E B system, an incremental conceptual clustering algorithm that uses an attribute-value language to represent its instances and concepts. The two systems are consequently similar in both structure and control. COBWEBR classifies instances described in a relational language, that is, each instance is a set of ground literals. COBWEBR represents concepts by storing an associated set of features, which correspond to the literals of the input language. The features are variabilized to allow them to record the relational structure of the instances without being bound to the particular objects being related. Associated with each feature is a conditional probability of occurrence, called the confirmed count, and a conditional probability of nonoccurrence, called the missing count. The concepts in the hierarchy are partially-ordered by generality. Each internal node contains all of the features of its children, storing them in such a way as to reflect the amount of overlap in their relational structures. If several of the children have a feature that can be consistently unified, that feature in the parent will have a high confirmed conditional probability.