Noise-Tolerant Conceptual Clustering

Fisher (1987a,b) introduced a performance task for conceptual clustering: flexible prediction of arbitrary attribute values, not simply the prediction of a single 'class' attribute. This paper extends earlier analysis by considering the effects of noise and other environmental factors. The degradation in flexible prediction accuracy that results from noise is mitigated by 'preferred' prediction points for individual attributes. Methods that identify these prediction points are inspired by pruning in learning from examples. We extend these noise-tolerant techniques to untutored learning. In addition, prediction point preferences shed light on relationships between conceptual clustering, case-based, and default reasoning.