A New Parallelization Method for K-means

Abstract: K-means is a popular clustering method used in data mining area. To work with large datasets, researchers propose PKMeans, which is a parallel k-means on MapReduce. However, the existing k-means parallelization methods including PKMeans have many limitations. PKMeans can't finish all its iterations in one MapReduce job, so it has to repeat cascading MapReduce jobs in a loop until convergence. On the most popular MapReduce platform, Hadoop, every MapReduce job introduces significant I/O overheads and extra execution time at stages of job start-up and shuffling. Even worse, it has been proved that in the worst case, k-means needs $2^{{\Omega}(n)}$ MapReduce jobs to converge, where n is the number of data instances, which means huge overheads for large datasets. Additionally, in PKMeans, at most one reducer can be assigned to and update each centroid, so PKMeans can only make use of limited number of parallel reducers. In this paper, we propose an improved parallel method for k-means, IPKMeans, which has a parallel preprocessing stage using k-d tree and can finish k-means in one single MapReduce job with much more reducers working in parallel and lower I/O overheads than PKMeans and has a fast post-processing stage generating the final result. In our method, both k-d tree and the new improved parallel k-means are implemented using MapReduce and tested on Hadoop. Our experiments show that with same dataset and initial centroids, our method has up to 2/3 lower I/O overheads and consumes less amount of time than PKMeans to get a very close clustering result.