Abstract/Summary

Clustering is an essential data mining technique that divides observations into groups where each group contains similar observations. K-Means is one of the most popular and widely used clustering algorithms that has been used for over fifty years. The majority of the running time in the original K-Means algorithm (known as Lloyd’s algorithm) is spent on computing distances from each data point to all cluster centres to find the closest centre to each data point. Due to the current exponential growth of the data, it became a necessity to improve KMeans even further to cope with large-scale datasets, known as Big Data. Hence, the main aim of this thesis is to improve the efficiency and scalability of Lloyd’s K-Means. One of the most efficient techniques to accelerate K-Means is to use triangle inequality. Implementing such efficient techniques on a reliable distributed model creates a powerful combination. This combination can lead to an efficient and highly scalable parallel version of K-Means that offers a practical solution to the problem of clustering Big Data. MapReduce, and its popular open-source implementation known as Hadoop, provides a distributed computing framework that efficiently stores, manages, and processes large-scale datasets over a large cluster of commodity machines. Many studies introduced a parallel implementation of Lloyd’s K-Means on Hadoop in order to improve the algorithm’s scalability. This research examines methods based on triangle inequality to achieve further improvements on the efficiency of the parallel Lloyd’s K-Means on Hadoop. Variants of K-Means that use triangle inequality usually require extra information, such as distance bounds and cluster assignments, from the previous iteration to work efficiently. This is a challenging task to achieve on Hadoop for two reasons: 1) Hadoop does not directly support iterative algorithms; and 2) Hadoop does not allow information to be exchanged between two consecutive iterations. Hence, two techniques are proposed to give Hadoop the ability to pass information from an iteration to the next. The first technique uses a data structure referred to as an Extended Vector (EV), that appends the extra information to the original data vector. The second technique stores the extra information on files where each file is referred to as a Bounds File (BF). To evaluate the two proposed techniques, two K-Means variants are implemented on Hadoop using the two techniques. Each variant is tested against variable number of clusters, dimensions, data points, and mappers. Furthermore, the performance of various implementations of K-Means on Hadoop and Spark is investigated. The results show a significant improvement on the efficiency of the new implementations compared to the Lloyd’s K-Means on Hadoop with real and artificial datasets.