面向网络结构发现的批量主动学习算法  

Batch Mode Active Learning for exploring structure of networks

作　　者：柴变芳 魏春丽 曹欣雨 王建岭 Chai Bianfang;Wei Chunli;Cao Xinyu;Wang Jianling(School of Information Engineering,Hebei GEO University,Shijiazhuang,050031,China;Office of Academic Affairs,Hebei University of Chinese Medicine,Shijiazhuang,050200,China)

机构地区：[1]河北地质大学信息工程学院,石家庄050031 [2]河北中医学院教务处,石家庄050200

出　　处：《南京大学学报（自然科学版）》2019年第6期1020-1029,共10页Journal of Nanjing University（Natural Science）

基　　金：国家自然基金(61503260,81473773);河北省自然科学基金(F2019403070);河北省人文社会科学项目(SD151087)

摘　　要：网络结构发现可识别网络多类型聚类模式,但其准确率有待提升.批量主动学习选择质量高的节点集合构造先验,可提升无监督网络结构发现的性能.面向属性网络分类的主动学习BMAL(Batch Mode Active Learning)只考虑链接信息实现网络节点选择,但不能有效选择使模型性能提升至最优的节点集合,且依赖初始人工标注及参数.提出一个新的批量主动学习算法,利用目标函数的子模性迭代选择最优的节点集合.该方法基于未标记节点的不确定性和非冗余影响力选择最优节点集合,不确定性依据节点及其邻居的类隶属度,影响力依据节点的非重叠中心性,两个指标的权重依据熵权法自动确定.人工和真实网络上的实验结果表明,该方法能选择使结构发现性能提升最大的节点集合.Exploring structure of networks can identify many types of clustering pattern,which helps people understand and utilize networks.However,its accuracy needs to be improved.Batch Mode Active Learning(BMAL)selects informative nodes to construct priors,which are used to improve the performance of unsupervised structure exploration.BMAL for classification on attributed networks can realize node selection for networks only based on links.But it cannot select optimal nodes which improve better performance of the model.In order to improve the performance of structure exploring,a new batch active learning algorithm is provided,which selects the optimal nodes iteratively on the submodular property of objective function.It selects a node set that maximizes the uncertainty and non redundant impact of unlabeled nodes.The uncertainty is computed according to the memberships of itself and its neighborhoods.And the impact is based on the nonoverlapping centrality.Their weights are adjusted by entropy weighted method automatically.Experiments on synthetic and real networks illustrate that our proposed method is able to select optimal node set that improves the performance of structure exploring.

关 键 词：批量主动学习 节点集合选择 网络结构发现 半监督聚类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]