All-Neighbor Classification Rule Based on Correlated Distance Combination

摘要

This report describes a new method of classifying data vectors by involving a two- step process. First, a data-specific step produces a distance qualitatively describing the similarity of the vector under analysis to each vector in a database representing a particular class. Second, the evidence represented by the vector of statistically correlated distances is combined into an overall numerical confidence that the vector under test belongs to the same class as the database vectors. In addition, the supporting evidence is available in the form of the individual distances. This all-neighbor method has several advantages over competing formalisms such as neural networks or the k-nearest-neighbor classification method. It can deal with data vectors of varying dimension, as long as the distance measure is capable of comparing them in some fashion. Even more importantly, it can deal with distance vectors of varying dimension, a common situation when dealing with a heterogeneous reference database. It produces a numeric confidence rather than just a simple classification. Further, it uses all the information contained in the distance vector, and it facilitates adjustment of false alarm rates. The method is applied to several different data types to demonstrate its generality.