KNN-DK: A Modified K-NN Classifier with Dynamic k Nearest Neighbors

Abstract

K-nearest neighbor (k-nn) is a widely used classifier in machine learning and data mining, and is very simple to implement. The k-nn classifier predicts the class label of an unknown object based on the majority of the computed class labels of its k nearest neighbors. The prediction accuracy of the k-nn classifier depends on the user input value of k and the distance measure used to compute the nearest neighbors from the training objects. If we use a static value for k for a particular classification task, the prediction accuracy of a k-nn classifier may decrease due to class imbalance in a dataset. In this paper, we propose a modified k-nn classifier that considers class imbalance in a dataset, and computes an appropriate value for k. The proposed k-nn classifier has been validated on a large number of benchmark datasets from various domains. The method is compared with traditional k-nn, decision tree, random forest and SVM classifiers, and the method yields significantly better prediction accuracy than the traditional the k-nn classifier and other algorithms.