Data decomposition for parallel K-means clustering

Developing fast algorithms for clustering has been an important area of research in data mining and other fields. K-means is one of the widely used clustering algorithms. In this work, we have developed and evaluated parallelization of k-means method for low-dimensional data on message passing computers. Three different data decomposition schemes and their impact on the pruning of distance calculations in tree-based k-means algorithm have been studied. Random pattern decomposition has good load balancing but fails to prune distance calculations effectively. Compact spatial decomposition of patterns based on space filling curves outperforms random pattern decomposition even though it has load imbalance problem. In both cases, parallel tree-based k-means clustering runs significantly faster than the traditional parallel k-means.