An Empirical Study of Spectral Social Network Partition on GPGPU Platforms

Social network analysis has attracted extensive research attention recently. By setting persons or more general entities as network vertices and using edges to represent their interactions, network provides an effective tool in representing complex social relations. Dividing such a network into different clusters in accordance with its inherent structure has been widely investigated. A method called modularity optimization is a principled and widely adopted way to handle this problem. But with high computational cost, an approximate solution, the spectral method, has been introduced to ensure the tractability for large-scale networks. Our work implemented the spectral method on a platform with general-purpose computing on graphics processing units (GPGPU) and compared the results with conventional solutions, which reported significantly improved running speed.