Lowering the complexity of k-means clustering by BFS-dijkstra method for graph computing

Anna Zhang, Jun Yao, Yasuhiko Nakashima

Abstract: K-means is a method of vector quantization, which is now popularly used for clustering analysis in massive data mining. Due to its heavily computational-intensive feature for iteratively re-computing and sorting distances, the execution of k-means takes a huge amount of time, especially when processing large graph data such as the practical social networks. This paper studies an alternative method to emulate the k-clustering from another view, in which the vertices in a graph are partitioned into k farthest clusters. This method can be implementable in a breadth-first-search (BFS) form and then becomes easily parallelizable. Our result shows that our BFS-based k-clustering achieves more than 100x speeds than the traditional partitioning in the open-source graphlab project.