当前位置: X-MOL 学术IEEE Trans. Netw. Sci. Eng. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Unraveling the Detectability of Stochastic Block Model With Overlapping Communities
IEEE Transactions on Network Science and Engineering ( IF 6.7 ) Pub Date : 2021-02-11 , DOI: 10.1109/tnse.2021.3058520
Huaying Wu , Luoyi Fu , Huan Long , Guie Meng , Xiaoying Gan , Yuanhao Wu , Haisong Zhang , Xinbing Wang

The detectability and distinguishability analysis for stochastic block model (SBM) have been very important topics in community detection. Detectability refers to the ability for weak recovery of communities in a graph, and distinguishability refers to the ability to distinguish the generating model of a given network from SBM and Erdos-Renyi (E-R) model. Numerous prior arts have focused on networks with non-overlapping communities. It has been proved that the threshold for detectability and distinguishability is equal to the Kesten-Stigum (K-S) threshold for networks with two communities and below the K-S threshold for networks with more than three communities. In this work, we revisit the detectability and distinguishability problems of community detection considering the overlapping characteristics. Specifically, given a graph generated from symmetric stochastic block model (SSBM) with overlapping communities, we are interested in whether we can achieve weak recovery on it and whether it is possible to distinguish it from a graph generated from E-R model. We studied these two problems for both 2-community case and q-community case, then the corresponding information-theoretic thresholds for the detectability of communities are derived. These results imply that as the strength of overlap increases, the thresholds become higher, thus community detection becomes harder. Therefore, we can conclude that the overlapping characteristics will weaken the detectability and distinguiability of communities in the network.

中文翻译:

揭示具有重叠社区的随机块模型的可检测性

随机块模型(SBM)的可检测性和可区分性分析一直是社区检测中非常重要的主题。可检测性是指对图中社区弱恢复的能力,可区分性是指将给定网络的生成模型与 SBM 和 Erdos-Renyi (ER) 模型区分开来的能力。许多现有技术都集中在具有非重叠社区的网络上。已经证明,对于具有两个社区的网络,可检测性和可区分性的阈值等于 Kesten-Stigum (KS) 阈值,而对于具有三个以上社区的网络,则低于 KS 阈值。在这项工作中,我们重新考虑了社区检测的重叠特征的可检测性和可区分性问题。具体来说,给定一个由具有重叠社区的对称随机块模型 (SSBM) 生成的图,我们感兴趣的是我们是否可以在其上实现弱恢复,以及是否可以将它与从 ER 模型生成的图区分开来。我们针对2-社区案例和q-社区案例研究了这两个问题,然后推导出了社区可检测性的相应信息理论阈值。这些结果意味着随着重叠强度的增加,阈值变得更高,因此社区检测变得更加困难。因此,我们可以得出结论,重叠特征会削弱网络中社区的可检测性和可区分性。我们感兴趣的是我们是否可以在其上实现弱恢复,以及是否可以将其与 ER 模型生成的图区分开来。我们针对2-社区案例和q-社区案例研究了这两个问题,然后推导出了社区可检测性的相应信息理论阈值。这些结果意味着随着重叠强度的增加,阈值变得更高,因此社区检测变得更加困难。因此,我们可以得出结论,重叠特征会削弱网络中社区的可检测性和可区分性。我们感兴趣的是我们是否可以在其上实现弱恢复,以及是否可以将其与 ER 模型生成的图区分开来。我们针对2-社区案例和q-社区案例研究了这两个问题,然后推导出了社区可检测性的相应信息理论阈值。这些结果意味着随着重叠强度的增加,阈值变得更高,因此社区检测变得更加困难。因此,我们可以得出结论,重叠特征会削弱网络中社区的可检测性和可区分性。然后推导出社区可检测性的相应信息理论阈值。这些结果意味着随着重叠强度的增加,阈值变得更高,因此社区检测变得更加困难。因此,我们可以得出结论,重叠特征会削弱网络中社区的可检测性和可区分性。然后推导出社区可检测性的相应信息理论阈值。这些结果意味着随着重叠强度的增加,阈值变得更高,因此社区检测变得更加困难。因此,我们可以得出结论,重叠特征会削弱网络中社区的可检测性和可区分性。
更新日期:2021-02-11
down
wechat
bug