A Computational Performance Study of Unsupervised Data Clustering Algorithms on GPU

TLDR: In this article , a comparative computational performance study between the sequential and the parallel versions of FCM, K-means and their 2 parallel versions is presented, focusing on the execution time of the parallel and sequential implementations in addition to the speed up of parallel version with respect to the sequential one.

Abstract: Classification task is a very popular preprocessing step in different research fields. Its main role is to separate the different components of an object or dataset into homogeneous regions or groups based on the similarity of properties and features. Among the most popular clustering algorithms we cite fuzzy C-means (FCM) and K-means. In this these iterative techniques, a distance metric between each actual dataset point and the estimated centroids is calculated at each iteration. In this paper, we implement four algorithms; sequential FCM, sequential K-means and their 2 parallel versions. A comparative computational performance study between the sequential and the parallel version is presented. This study, will focus on the execution time of the parallel and sequential implementations in addition to the speed up of the parallel version with respect to the sequential one. The experimental tests were conducted on a randomly generated number dataset. The parallel versions were implemented on a SIMD architecture of Nvidia GPU. The influence of the variation of the data size and the number of clusters on the execution time was analyzed and interpreted.