Selecting Typical Instances in Instance-Based Learning

Abstract

Concepts involved in real world applications usually possess graded structures. Instead of being equivalent, instances of a concept may be characterized by a degree of typicality in representing the concept. Typical instances of a concept usually better characterize the concept than atypical instances do. This paper presents an instance-based learning approach in which typical instances are selected to store as concept descriptions. It first addresses the issue of measuring the typicality of an instance with respect to its concept. Then, it empirically shows that some concepts in standard datasets do have graded structures. Finally, it presents a simple instance-based learning and classification algorithm that successfully uses typicalities of instances. This approach has been tested on both artificial and practical domains, and compared with three different IBL approaches. The experimental results showed that the approach recorded lower storage requirements and higher classification accuracies than previous instance-based algorithms on several domains.