点击次数：

论文类型：期刊论文

发表时间：2016-05-01

发表刊物：INTERNATIONAL JOURNAL OF MODERN PHYSICS C

收录刊物：SCIE

卷号：27

期号：5

ISSN号：0129-1831

关键字：Complex networks; network measurements; nonlinear redundancy; nonlinear dimensionality reduction; Bayesian classifier

摘要：Discriminating complex networks is a particularly important task for the purpose of the systematic study of networks. In order to discriminate unknown networks exactly, a large set of network measurements are needed to be taken into account for comprehensively considering network properties. However, as we demonstrate in this paper, these measurements are nonlinear correlated with each other in general, resulting in a wide variety of redundant measurements which unintentionally explain the same aspects of network properties. To solve this problem, we adopt supervised nonlinear dimensionality reduction (INDR) to eliminate the nonlinear redundancy and visualize networks in a low-dimensional projection space. Though unsupervised NDR can achieve the same aim, we illustrate that supervised NDR is more appropriate than unsupervised NDR for discrimination task. After that, we perform Bayesian classifier (IBC) in the projection space to discriminate the unknown network by considering the projection score vectors as the input of the classifier. We also demonstrate the feasibility and effectivity of this proposed method in six extensive research real networks, ranging from technological to social or biological. Moreover, the effectiveness and advantage of the proposed method is proved by the contrast experiments with the existing method.