Batch Mode Active Learning For Exploring Structure Of Networks

With the progress of industrialization,urbanization and the rise of new communication technologies,a large number of various online social networking platforms have emerged,such as Friendster,Myspace,Bebo,Facebook,weibo and tieba,etc.It has become one of the current research hotspots to discover the clustering structure and interaction rules of these networks.Researchers have proposed a large number of unsupervised network clustering structure discovery methods,but their clustering accuracy is not high.Semi-supervised network clustering can use prior information to improve the accuracy of network clustering,but it depends on the quality of prior information.Active learning can select high-quality network nodes to mark,which not only guarantees the quality of prior information,but also ensures the maximum clustering performance improvement with less priors.Therefore,active learning oriented to network structure discovery has important research value.At present,some researchers have proposed some active learning methods for community discovery,but the selected network nodes have no effect on the network structure with mixed mode.A network with mixed modes may not have a community structure,or may have other clustering structures,such as binary structures,star structures,and mixtures of multiple structures.Therefore,it is necessary to design an active learning algorithm to improve the performance of network structure discovery with multiple clustering patterns.This paper mainly completes the following research contents:(1)Aiming at the problem that the Batch Mode Active Learning algorithm of attributionoriented network classification(BMAL)cannot effectively select the optimal set of nodes that can improve the performance of network structure discovery to the maximum by only considering the network node link information,this paper proposes a Batch Mode Active Learning algorithm BMAL_NMS(Network with Mixture Structures)oriented network structure discovery.The algorithm selects the optimal set of nodes based on the three strategies of maximum uncertainty,maximum influence and minimum redundancy of unmarked nodes.The experimental results show that BMAL_NMS algorithm can select the node set that can improve the performance of network structure.(2)Aiming at the problem that BMAL_NMS algorithm cannot make full use of network node attribute information,it integrates network node attribute information and link information,and proposes a batch active learning algorithm BMAL_CLF(Content and Link Fusion)oriented to attribute network.The algorithm uses representation learning to represent node attribute information and link information as node vectors,then takes node vectors as input of the algorithm,and selects the optimal set of nodes iteratively by using the sub-model of the objective function.Experimental results show that BMAL_CLF algorithm is superior to BMAL_NMS algorithm.(3)BMAL_CLF,a batch active learning algorithm oriented to attribute network,is applied to the portrait of CSDN users.According to various behavioral data of users and blog content data published,representative CSDN users are selected.The selected user is taken as prior information and the semi-supervised attribute network structure discovery algorithm is run.The experimental results show that BMAL_CLF has certain application value.