Title: An influence-based k-nearest neighbour classifier for classification of data with different densities

Authors: Hassan Motallebi; Amir-Hossein Fakhteh

Addresses: Faculty of Electrical and Computer Engineering, Graduate University of Advanced Technology, Kerman, Iran ' Faculty of Electrical and Computer Engineering, Graduate University of Advanced Technology, Kerman, Iran

Abstract: 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.

Keywords: local mean-based KNN; modified distance measure; reverse KNN; gradient descent direction; different densities.

DOI: 10.1504/IJBIDM.2024.140241

International Journal of Business Intelligence and Data Mining, 2024 Vol.25 No.2, pp.147 - 167

Received: 31 Jan 2022
Accepted: 05 Dec 2022
Published online: 31 Jul 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article