Do We Need Whatever More Than k-NN?

Many sophisticated classification algorithms have been proposed. However, there is no clear methodology of comparing the results among different methods. According to our experiments on the popular datasets, k-NN with properly tuned parameters performs on average best. Tuning the parametres include the proper k, proper distance measure and proper weighing functions. k-NN has a zero training time and the test time can be significantly reduced by prior reference vector selection, which needs to be done only once or by applying advanced nearest neighbor search strategies (like KDtree algorithm). Thus we propose that instead of comparing new algorithms with an author's choice of old ones (which may be especially selected in favour of his method), the new method would be rather compared first with properly tuned k-NN as a gold standard. And based on the comparison the author of the new method would have to aswer the question: "Do we really need this method since we already have k-NN?"