Mux-Kmeans: multiplex kmeans for clustering large-scale data set

Kmeans clustering algorithm is widely used in a number of scientific applications due to its simple iterative nature and ease of implementation. The quality of clustering result highly depends on the selection of initial centroids. Different selections of initial centroids result in different clustering results. In practice, people run a series of Kmeans processes with multiple initial centroid groups serially and return the best clustering result among them. However, in the era of big data, a Kmeans process is implemented on MapReduce to scale to large data sets. Even a single Kmeans process on MapReduce requires considerable long runtime. This paper proposes Mux-Kmeans. Rather than executing multiple Kmeans processes serially, Mux-Kmeans launches these Kmeans processes concurrently with multiple centroid groups. In each iteration, Mux-Kmeans (i) evaluates these Kmeans processes, (ii) prunes the low-quality Kmeans processes, and (iii) incubates new Kmeans processes. After a certain number of iterations, it finally obtains the best among these local optimal results. We implement Mux-Kmeans on MapReduce and evaluate it on Amazon EC2. The experimental results show that starting from the same initial centroid groups, the clustering result of Mux-Kmeans is always non-worse than the best of a series of Kmeans processes. Mux-Kmeans also saves elapsed time than serial multiple Kmeans processes.