On the Missing Link between Frequent Pattern Discovery and Concept Formation

Concept Formation is a unsupervised learning task usually decomposed into the two subtasks of clustering and characterization. This paper presents a novel approach to Concept Formation in First Order Logic (FOL) which adopts a pattern-based approach to clustering and a bias-based approach to characterization. The resulting method extends therefore the levelwise search method for Frequent Pattern Discovery. The FOL fragment chosen is Aℒ-log, a hybrid language that merges the description logic AℒC and the clausal logic DATALOG and turns out to be suitable for applications in the context of Ontology Refinement. Indeed the method returns a taxonomy rooted into the concept that occurs in an existing taxonomie ontology and needs to be refined in the light of new knowledge coming from an external data source. Experimental results have been obtained on an AℒC ontology enriched with DATALOG data extracted from the on-line CIA World Fact Book.