Title: Particle swarm optimisation-based KNN for improving KNN and ensemble classification performance

Authors: Debojit Boro; Dhruba K. Bhattacharyya

Addresses: Department of Computer Science and Engineering, Tezpur University, Tezpur, Assam, Pin – 784028, India ' Department of Computer Science and Engineering, Tezpur University, Tezpur, Assam, Pin – 784028, India

Abstract: In this paper we use KNN algorithm for our ensemble classification process that finds out the nearest K training samples given a test sample where each sample is predictions vector generated by combined computation of classifiers ensemble and algebraic combiners. In this attempt to reduce the computational time involved in finding K training samples by KNN, we used particle swarm optimisation (PSO) with KNN which randomly selects the training samples from the training set until a global consensus is reached among the particles and label that test sample with an appropriate class by weighted majority voting (WMV) of K training samples. The proposed method demonstrated better ensemble performance as compared to KNN and other traditional ensemble methods in terms of computational time and generalisation accuracies when tested over several datasets from UCI repository and other high dimensional datasets.

Keywords: classifiers; ensemble classification; decision tree; naive Bayes; decision table; random forests; algebraic combiners; similarity measures; term frequency; inverse document frequency; KNN; k-nearest neighbour; particle swarm optimisation; PSO; evolutionary computation; ensemble performance.

DOI: 10.1504/IJICA.2015.073004

International Journal of Innovative Computing and Applications, 2015 Vol.6 No.3/4, pp.145 - 162

Received: 07 Feb 2015
Accepted: 14 Jul 2015
Published online: 11 Nov 2015 *

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