Voting over multiple k-NN classifiers

One issue for the popular k-nearest neighbor decision rule concerns the choice of the number of neighbors k. We present a novel approach to k-NN, namely the classification is performed with an ensemble of k-NN classifiers, each trained on a random partition of the whole training set and thus having its own k. As opposed to most ensemble schemes, the classification speed in our algorithm is on par with the speed of original k-NN. The effectiveness of the proposed algorithm is confirmed on a quality control application task.