CD-KNN: A Modified K-Nearest Neighbor Classifier with Dynamic K Value

Abstract

K-Nearest Neighbor (KNN) is a widely used classifier in many applications of Machine Learning. KNN is a distance based learning that uses different distance measures to compute the K nearest neighbors of a test instance. But, it is very challenging to pick up an appropriate K value for KNN and especially, for an imbalance dataset, the value of K plays a major role during classification of unknown instances by KNN. So, the paper addresses the issue of computing the K value dynamically for each test instance to be classified by the KNN classifier. We developed a modified KNN classifier called Cluster-based Dynamic KNN (CD-KNN) that computes the K nearest neighbors dynamically and the performance of the classifier is evaluated using 9 datasets. From the experimental results, we observed that the proposed KNN classifier yields better results as compared to the traditional classifier on various datasets.