Parallel Evolving Clustering Method for Big Data Analytics Using Apache Spark: Applications to Banking and Physics

摘要

A novel parallel implementation of the Evolving Clustering Method (ECM) is proposed in this paper. The original serial version of the ECM is the clustering method which computes online and with a single-pass. The parallel version (Parallel ECM or PECM) is implemented in the Apache Spark framework, which makes it work in real time. The parallelization of the algorithm aims to handle a dataset with large volume. Many of the extant clustering algorithms do not involve a parallel one-pass method. The proposed method addresses this shortcoming. Its effectiveness is demonstrated on a credit card fraud dataset (with size 297 MB), and a Higgs dataset was taken from Physics pertaining to particle detectors in the accelerator (with size 1.4 GB). The experimental setup included a cluster of 10 machines having 32 GB RAM each with Hadoop Distributed File System (HDFS) and Spark computational environment. A remarkable achievement of this research is a dramatic reduction in computational time compared to the serial version of the ECM. In future, the PECM shall be hybridized with other machine learning algorithms for solving large-scale regression and classification problems.