The multi-class metric problem in nearest neighbour discrimination rules

摘要

Nearest neighbour rules classify a new data point on the basis of the class of the closest design set point. A prerequisite for this is a choice of metric by which distance may be calculated. Metrics which have been suggested for the two-class case fall into two classes: global and local. In this paper local metrics are described and are generalized to the multi-class case. Properties of various possible generalizations are compared, and the results of a simulation study are described.