大规模网络广义社区发现随机变分推理算法

Stochastic variational inference algorithms for massive networks with general structures

流行度-生成度随机块(popularity-productivity stochastic block,PPSB)模型能发现网络广义社区,但该模型易过拟合,且不能有效处理大规模网络,故提出一个3层贝叶斯网络广义社区发现(generalized PPSB,GPPSB)模型,并给出实现大规模链接网络和内容网络广义社区发现的随机变分推理(stochastic variational inference,SVI)算法GPPSB-SVI和GPPSB-C-SVI。不同规模人工网络和实际网络上的实验结果表明:GPPSB-SVI准确性优于已有流行大规模网络社区发现算法,效率高于基于PPSB模型的广义社区发现算法;GPPSB-C-SVI准确性优于GPPSB-SVI算法;GPPSB模型引入节点隶属度和类间链接概率矩阵的先验分布,可更好地对网络建模,其参数估计算法GPPSB-SVI、GPPSB-C-SVI可更有效地实现大规模网络广义社区发现。

PPSB model could explore general structure of a network, but it was easy to over fit and unable to deal with large net- works. A three-level hierarchical Bayesian model GPPSB was presented. Stochastic variational inference algorithms for massive networks based on links （ GPPSB - SVI） and based on contents and links （ GPPSB - C - SVI） were inferenced. Experiments on synthetic and real networks show that GPPSB -SVI is more accuracy than existing popular community detection algorithms on massive networks and more efficiency than general community detection algorithms based on the PPSB model, and GPPSB - C - SVI is superior to GPPSB - SVI. GPPSB draws the prior generative models of memberships of nodes and link probabilities between clusters for better modeling, in which GPPSB -SVI and GPPSB -C -SVI are able to deal with massive networks more efficiently.