An Empirical Method to Improve Edit-distance Parameters for a Nearest-neighbor-based Classiication Task

This paper describes a method for automatically improving the parameters of the edit distance used in a given nearest-neighbor pattern classiication task. The method uses a simple self-consistent procedure which assigns costs to the basic editing operations according to their discriminative power. The discriminative powers of operations are deened based on the statistics of their use when a prototype set is used to classify itself. The parameters obtained are not optimal but may be obtained easily and the experimental evidence |a handwritten numeral classiication task| shows that the performance of the classiier may be improved easily by using the proposed method on the prototype set.