KNN Kernel Shift Clustering with Highly Effective Memory Usage

This paper presents a novel clustering algorithm with highly effective memory usage. The algorithm, called kNN kernel shift, classifies samples based on underlying probability density function. In clustering algorithms based on density, a local mode of the density represents a cluster center. It is effective to shift each sample to a point having higher density, considering the density gradient. Estimation of density and determination of the shifting are calculated using distance between samples. Large memory is necessary because the number of all combinations of N samples is O(N 2 ). We propose a mode seeking approach using only neighbor samples in order to save memory. Experimental results show the effectiveness of the proposed algorithm in terms of clustering accuracy, processing time, and memory usage.