Grouping game players using parallelized K-Means on supercomputers

Grouping game players based on their online behaviors has attracted a lot of attention recently. However, due to the huge volume and extreme complexity in online game data collections, grouping players is a challenging task. This study has applied parallelized K-Means on Gordon, a supercomputer hosted at San Diego Supercomputer Center, to meet the computational challenge on this task. By using the parallelization functions supported by R, this study was able to cluster 120,000 game players into eight non-overlapping groups and speed up the clustering process by one to four times under the two- To eight-degree of parallelization. This study has systematically examined a number of factors which may affect the quality of the clusters and/or the performance of the clustering processes; those factors include degree of parallelism, number of clusters, data dimensions, and variable combinations. This study invented a method to identify the optimal clustering schema, which can choose the most discriminative features and create an appropriate number of clusters in K-Means clustering. Besides demonstrating the effectiveness of parallelized K-Means in grouping game players, this study also highlights some lessons learned for using K-Means on very large datasets and some experience on applying parallel processing techniques in intensive data analysis.