An evidence-theoretic k-NN rule with parameter optimization

摘要

The paper presents a learning procedure for optimizing thenparameters in the evidence-theoretic k-nearest neighbor rule, a patternnclassification method based on the Dempster-Shafer theory of beliefnfunctions. In this approach, each neighbor of a pattern to be classifiednis considered as an item of evidence supporting certain hypothesesnconcerning the class membership of that pattern. Based on this evidence,nbasic belief masses are assigned to each subset of the set of classes.nSuch masses are obtained for each of the k-nearest neighbors of thenpattern under consideration and aggregated using Dempster's rule ofncombination. In many situations, this method was found experimentally tonyield lower error rates than other methods using the same information.nHowever, the problem of tuning the parameters of the classification rulenwas so far unresolved. The authors determine optimal or near-optimalnparameter values from the data by minimizing an error function. Thisnrefinement of the original method is shown experimentally to result innsubstantial improvement of classification accuracy