We consider the community recovery problem on a one-dimensional random geometric graph where every node has two independent labels: an observed location label and a hidden community label. A geometric kernel maps the locations of pairs of nodes to probabilities. Edges are drawn between pairs of nodes based on their communities and the value of the kernel corresponding to the respective node locations. Given the graph so generated along with the location labels, the latent communities of the nodes are to be inferred. In this work, we will look into the fundamental statistical limits for recovering the communities in such models. Additionally, we propose a linear-time algorithm (in the number of edges) and show that it recovers the communities of nodes exactly up to the information theoretic threshold.
Community detection in the sparse hypergraph stochastic block model
University of Washington(华盛顿大学);University of California, San Diego(加州大学圣地亚哥分校)
本文研究了稀疏随机超图中的社团检测问题。Angelini et al.(2015)证明了在由超图随机块模型生成的稀疏超图中,社区检测的模型参数存在一个尖锐的阈值。我们解决了两个块的情况下的猜想的积极的一部分:在阈值以上,有一个谱算法,渐近几乎肯定构建一个分区的超图与真正的分区。我们的方法是Massoulié(2014)为稀疏随机图开发的方法的随机超图的推广。
Active Learning for Community Detection in Stochastic Block Models
University of Southern California(南加州大学)
随机块模型(SBM)是网络科学和机器学习中随机图的一个重要生成模型,可用于基准测试社区发现(或聚类)算法。对称SBM生成具有$2n$节点的图,这些节点聚类成两个相等大小的社区。节点在社区内以概率$p$连接,在不同社区之间以概率$q$连接。我们考虑$p=a\ln (n)/n$和$q=b\ln (n)/n$的情况。在这种情况下,最近表明,以高概率(w. h. p.)仅使用该图是可能的,当且仅当B-H(CH)发散$D(a,b)=(\sqrt{a}-\sqrt{b})^2 \geq 1$。在这项工作中,我们研究了社区发现查询有限数量的所选节点的标签(即,主动学习)。我们的主要结果是表明,在一定条件下,采样标签的一个消失的小部分节点(一个数字次线性$n$)是足够的精确社区发现,即使当$D(a,b)<1$。此外,我们提供了一个有效的学习算法,恢复社区成员的所有节点w. h. p.,只要采样点的数量满足充分条件。我们还表明,恢复是不可能的,如果观察到的标签的数量小于$n^{1-D(a,b)}$。通过数值实验证明了我们的结果的有效性。
Sharon High School(沙伦高中);Network Science Institute, Northeastern University(东北大学网络科学研究所);Complexity Science Hub, Vienna(维也纳复杂性科学中心);Department of Physics, Northeastern University(东北大学物理系);Department of Mathematics, Northeastern University(东北大学数学系);Department of Electrical & Computer Engineering, Northeastern University(东北大学电气与计算机工程系)