An adaptive graph sampling framework for graph analytics

In large-scale data processing, graph analytics of complex interaction networks are indispensable. As the whole graph processing and analytics can be inefficient and usually impractical, graph sampling by keeping a portion of the original graph becomes a favorable approach. While prior work focused on fixed edge and node selection strategy based on predetermined criteria, without adaptive feedback to adjust the sampling process, this type of sampling algorithms has limited flexibility and estimation accuracy for complex graphs. In this paper, we propose an adaptive graph sampling framework, and design AdapES, an adaptive edge sampling algorithm based on this framework. Compared to non-adaptive sampling methods, our approach can continually monitor the difference between the current sampled subgraph and the original graph, and dynamically adjust the edge sampling probability based on this observed sampling difference. Guided by a preset sampling goal, this algorithm automatically adapts to the fluctuations in the random sampling process with high flexibility. The experimental evaluation in 11 datasets demonstrates that AdapES outperforms other algorithms for preserving various graph properties and statistics.