Nearest Neighbor Rules for Statistical Classification Based on Ranks.

摘要

Some rank-based rules for classifying an observation into one of two populations, from which training samples are available, are suggested. The limiting values of their probabilities of misclassification are derived as the sizes of the training samples tend to infinity. The nearest neighbor idea is used in terms of ranks and two-sided neighbors to define the rank nearest neighbor rule. A multi-stage version is also suggested (in case of ties) and studied. (Author)