A new approach to determine the parameters of dissimilarity function for the evidence-theoretic k-NN classification rule

This paper presents a new approach to determine the parameters in the evidence-theoretic k-NN classification rule. Given a pattern recognition problem, we first compute a reference nearest neighbor distance to separate samples of one class from other samples with least error rate, and then calculate parameters of dissimilarity measure function based on it. Under the condition of small scale samples with nonGaussian distribution, the proposed method can get more suitable parameters and thus reduce classification error rate. And its computation complexity is 4-8 times lower than that of L.M. Zouhal´s method.