Predicting the number of nearest neighbor for kNN classifier

The k nearest neighbor (kNN) rule is known as its simplicity, effectiveness, intuitiveness and competitive classification performance. Selecting the parameter k with the highest classification accuracy is crucial for kNN. There's no doubt that the leave-one-out cross validation (LOO-CV) is the best method to do this work as its almost unbiased property. However, it is too time consuming to be used in practice especially for large data. In this paper, we propose a new algorithm for selecting an optimal neighborhood size k. We found that the classification accuracy of LOO-CV is approximate concave for the parameter k. And a search method is proposed to pick out the optimal value of k. An empirical study conducted on 8 standard databases from the UCI repository shows that the new strategy can find the optimal k with significantly less time than the LOO-CV method. © 2019 International Association of Engineers.