众大咖会师复网大会探寻关键节点挖掘

转载

DataCastle数据城堡 2021-07-28 16:07:01

文章标签 复网大会 文章分类 代码人生

DataCastle国际大师赛完美收官，一起来围观第十三届全国复杂网络大会“网络挖掘”分会场吧~

11月26日上午，深圳大学文科楼H07

这里是第十三届全国复杂网络大会“网络挖掘”分会场，同时也是DataCastle国际大师赛收官活动——大赛优秀成果报告会的现场。前来听取报告的是来自全国各地的复杂网络领域学者及研究人员，他们强势围观并站满了报告厅。

活动现场，大师赛中ABM代表中山大学胡延庆、中国科学院理论物理研究所周海军、电子科技大学吕琳媛以及优秀参赛队伍代表胡清华、朱冯贶天、任涛，特邀嘉宾中国科学技术大学李明、电子科技大学郝东分别为到场的两百多位来宾分享了他们的报告。

报告概要一览

本次比赛邀请了国内外复杂网络领域8名顶级专家学者作为大赛的Academic Board Member，担任大赛评委并为大赛提供学术指导。这次，他们中的三位莅临复网大会成果汇报现场并带来精彩报告。

胡延庆

报告人：胡延庆（中山大学）

题目：复杂网络边的可预测性理论及其价值

摘要：From presidential election toviral marketing even for containment of terrorism, the optimal monitoring andpromoting the diffusion of information in large-scale online social networks(OSNs) have great values. However, the OSNs have gigantic sizes and their fullstructural data are usually unavailable, let alone these problems are NP-hard.Therefore, identifying the monitoring nodes and optimal spreaders of the wholenetwork are commonly believed as two big challenges. By mapping the spreadingdynamics onto bond percolation in statistical physics, we show these twodifferent problems can be reduced to one problem. Moreover, we find that inmany stochastic spreading events, the information spreading happens in only oneof two well separated phases: a locally confined phase and a global viralspreading phase. This motivates an efficient algorithm with constant timecomplexity on the long-standing problems of best seed spreaders and monitoringnodes selection, whose performance remarkably close to the true optimum.

周海军

报告人：周海军（中国科学院理论物理研究所）

题目：圈环的统计物理学及应用

摘要：圈(loop) 是无向网络中的回路结构，而环 (cycle) 则指有向网络中的有向回路结构。圈环是网络的一种非常重要的结构特征，他们的存在导致网络中的动力学过程有反馈相互作用，从而带来动力学复杂性；圈环的存在也是网络上许多组合优化问题由容易变为困难的关键原因之一。在该报告中，我将介绍近年来利用自旋玻璃统计物理方法研究复杂网络圈环性质的一些工作，尤其侧重于介绍在无向网络最小反馈节点集合问题 [1] 和有向网络最小反馈边集合问题 [2] 上取得的进展，以及这些进展对于挖掘网络中关键节点 [3] 和关键边 [4] 的可能应用。

附：

[1] Hai-Jun Zhou, Spin glass approachto the feedback vertex set problem. European Physical Journal B 86:455 (2013).

[2] Yi-Zhi Xu, Hai-Jun Zhou, OptimalSegmentation of directed graph and the minimum number of feedbackarcs. Journal of Statistical Physics 169: 187–202 (2017).

[3] Salomon Mugisha, Hai-Jun Zhou,Identifying optimal targets of network attack by belief propagation. PhysicalReview E 94: 012305 (2016).

[4] Jin-Hua Zhao, Hai-Jun Zhou, Feedbackarcs and node hierarchy in directed networks. Chinese Physics B 26: 078901(2017).

吕琳媛

报告人：吕琳媛（电子科技大学、阿里巴巴）

题目：H指数及其衍生指标

摘要：一直以来，学术影响力评价都是一个备受关注的话题。主要原因在于其所涉及的应用领域极为广泛。从论文的价值衡量到期刊的影响力刻画，从科研人员的业绩考核与晋升到诺贝尔奖的评选，从双一流高校的排名到国家创新能力的评估，等等，这些问题都涉及如何对研究成果的学术影响力进行科学的评价。在众多的指标中，最著名的非H-指数莫属。一个学者的H-指数等于h表示在其所有论文中，至多有h篇论文每篇被引用至少h次。H-指数自2005年提出以来，学者们对它的“热情”至今丝毫未减。到目前为止已经有近三十多种H-指数的衍生指标，本报告将选取具有代表性的一些指标进行介绍，并对他们各自的特点和优劣势进行比较分析，最后给出一些应用示例。

此次大师赛总计吸引了全球各地759名选手参赛，根据选手的行程安排，我们邀请了660支队伍中的三支优秀队伍带来他们的参赛成果汇报。

胡清华

报告人：胡清华（格式不正确电子科技大学）

题目：关于图分割的关键节点挖掘算法

摘要：关于图分割的关键节点挖掘算法采用鲁棒性评价指标，基本渗流目标是使得渗流过程中网络的连通分支大小尽可能小。不使用传统的节点数据指标，例如度数，介数，核数，h family等数据指标，采用了新的节点重要性衡量方式，动态的更新节点的重要性指标。在此基础上，可以采用递归或者贪心的策略，从质量和速度两个不同的侧重角度求得网络的渗流结果。

朱冯贶天

报告人：朱冯贶天（zhfkt 个人）

题目：Identifying vital nodes based on Collective Influence and ImprovedReinsertaion

摘要：The presentation will introduce the solution ofalgorithm in identifying vital nodes in complex networks for Datacastle MasterCompetition , including Collective Influence and Improved Reinsertion. Comparedwith traditional algorithms including Betweenness, Closeness, PageRank, Degreeand Collective Influence, the new proposed algorithm of reconstructingcollapsed complex network to find most influential nodes is able to achieve thebetter performance on the 8 competition datasets ,provided by DataCastleMaster Competition, in terms of robustness and speed.

任涛

报告人：任涛（白山黑水东北大学）

题目：重要节点排序算法

摘要：大师赛评分标准的最优结果是按照节点排序算法得出的节点顺序去除节点，使每次去除节点后的剩余网络的最大连通子图的累加和最小。我们的算法从该目标着手，采用逆向思维，从不重要的节点找起，并基于贪婪策略，每次迭代选择使最大连通子图增长最小的节点，进而得到节点排序结果。

主办方还邀请了在复杂网络挖掘领域颇有心得的两位学者进行特邀嘉宾报告。

李明

报告人：李明（中国科学技术大学）

题目：Identify structural nodes by percolation: The non-uniquesolutions

摘要：Identifying the structure of complex networks is one of theimportant issues in network science, including community detection, noderanking, network robustness, and so on. In one word, identifying the structureof networks is nothing but dividing nodes into groups based on their topologiesor functions. As a classical model in statistical physics, percolation is justa theory that considers the emergence of connected clusters in networkedsystems. So it is very straightforward to introduce the percolation theory intothe identification of structural nodes. When we apply a percolation process ona network, the cluster forming of each node depends on the mesoscale connectionof nodes (strictly speaking, depending on the occupied probability), neitherlocal nor global connections. In light of this, we can expect to find a generalmethod for vital node identification, which can perform better on any networks.Furthermore, due to the randomness in the percolation model, we can finddifferent node rankings in different realizations (sometimes, for somenetworks, almost totally different rankings). However, these different rankingscan give almost the same performance in the test of node ranking, such asrobustness test. This indicates that the importance of a node is also dependenton the states of other nodes, and the solution of node ranking is not unique.

郝东

报告人：郝东（电子科技大学）

题目：Link Prediction via unified linear optimization

摘要：Link prediction is an elemental challenge in network science, whichhas already found applications in guiding laboratorial experiments, digging outdrug targets, recommending friends in social networks, probing mechanisms innetwork evolution, and so on. With a simple assumption that the likelihood ofthe existence of a link between two nodes can be unfolded by a linear summationof neighboring nodes' contributions, we obtain the analytical solution of theoptimal likelihood matrix, which shows remarkably better performance inpredicting missing links than the state-of-the-art algorithms for not onlysimple networks, but also weighted and directed networks. To our surprise, evensome degenerated local similarity indices from the solution outperformwell-known local indices, which largely refines our knowledge, for example, thedirect count of the number of 3-hop paths between two nodes more accuratelypredict missing links than the number of 2-hop paths (i.e., the number ofcommon neighbors), while in the previous studies, as indicated by the localpath index and Katz index, the statistics on longer paths are always consideredto be complementary to but less important than those on shorter paths.