Search

The k-nearest neighbour (KNN) is a simple and yet effective classification rule. To achieve robustness against outliers, several local mean-based extensions of the KNN classifier have been proposed which assign the query to the class with nearest local mean. However, using the conventional proximity measures causes poor performance in situations with multi-scale classes. Here, we propose a new local mean-based KNN classifier that uses a new modified distance measure which adjusts the proximity around each sample with respect to the situation and density. We scale the distance from the sample to each class with respect to the size of its influence set such that the sample seems closer to classes in which it has a higher number of reverse neighbours. We apply the proposed method on three local mean-based KNN classifiers. Our results show that the proposed method improves the performance of the local mean-based classifiers.