An Adaptive Hybrid and Cluster-Based Model for Speeding Up the k-NN Classifier

Abstract

A well known classification method is the k-Nearest Neighbors (k-NN) classifier. However, sequentially searching for the nearest neighbors in large datasets downgrades its performance because of the high computational cost involved. This paper proposes a cluster-based classification model for speeding up the k-NN classifier. The model aims to reduce the cost as much as possible and to maintain the classification accuracy at a high level. It consists of a simple data structure and a hybrid, adaptive algorithm that accesses this structure. Initially, a preprocessing clustering procedure builds the data structure. Then, the proposed algorithm, based on user-defined acceptance criteria, attempts to classify an incoming item using the nearest cluster centroids. Upon failure, the incoming item is classified by searching for the k nearest neighbors within specific clusters. The proposed approach was tested on five real life datasets. The results show that it can be used either to achieve a high accuracy with gains in cost or to reduce the cost at a minimum level with slightly lower accuracy.