Donald Bren School of Information and Computer Sciences

The goal of our research is to understand the power and appropriateness of instance-based representations and their associated acquisition methods. This paper concerns two methods for reducing storage requirements for instance-based learning algorithms. The first method, termed instance-saving, represents concept descriptions by selecting and storing a representative subset of the given training instances. We provide an analysis for instance-saving techniques and specify one general class of concepts that instance-saving algorithms are capable of learning. The second method, termed instance-averaging, represents concept descriptions by averaging together some training instances while simply saving others. We describe why analyses for instance-averaging algorithms are difficult to produce. Our empirical results indicate that storage requirements for these two methods are roughly equivalent. We outline the assumptions of instance-averaging algorithms and describe how their violation might degrade performance. To mitigate the effects of non-convex concepts, a dynamic thresholding technique is introduced and applied in both the averaging and non-averaging learning algorithms. Thresholding increases the storage requirements but also increases the quality of the resulting concept descriptions.