Learning Despite Concept Variation by Finding Structure

Learning accuracy depends on concept variation. The accuracy of six learning systems (C4.5, Grove, Greedy, Fringe, LFC and MRP) is compared using a set of forty test concepts. The selection of these concepts was guided by the existence of structured concepts that appear in diicult real-world domains (such as protein folding). Such concepts often have embedded, implicit structure , which may be revealed through explicit relations. Experiments using these benchmark concepts show that concept variation aaects systems that nd relations, such as MRP, signiicantly less than the other learners. We analyze this distinctive behavior in terms of concept characteristics and relate it to system performance in real-world domains.