融合结构和聚类的对称非负矩阵分解链路预测

Link prediction based on symmetric nonnegative matrix factorizationcombining structure and clustering

大部分链路预测算法仅单一考虑节点聚类或链接聚类而忽略网络结构与聚类内在关联性导致预测准确度下降.针对此问题,提出基于对称非负矩阵分解(SNMF)链路预测框架融合多类型结构和聚类信息捕获网络保持网络局部、全局以及节点和链接聚类.首先,融合节点和链接聚类系数(NEC)捕获节点邻域相关联程度,再将无向无权3个基于局部相似度方法共同邻居(CN)、资源分配(RA)和Adamic-Adar(AA)与聚类相融合同时保持结构和聚类;其次,将邻接矩阵映射到低维潜在空间,利用图正则化融合以上信息分别提出3个链路预测模型即SNMF-NEC-CN、SNMF-NEC-AA和SNMF-NEC-RA;此外,通过迭代更新规则学习所提模型参数,获得最优预测概率矩阵.在6个网络上与现有代表性方法比较,实验结果显示所提模型AUC和F1值分别提高了22%和11.4%.

Most of the existing link prediction algorithms only consider node clustering or link clustering and ignore the internal correlation between network structure and clustering,which leads to a decrease in prediction accuracy.In view of the above shortcomings,we propose a link prediction framework based on symmetric non⁃negative matrix fac⁃torization(SNMF)to fuse structure and cluster information capture network to maintain network local,global and node and link clustering.Firstly,the fusion node and edge clustering coefficient(NEC)captures the degree of node neighborhood association,and then undirected and weighted three local similarity methods Common Neighbor(CN),Resource Allocation(RA)and Adamic⁃Adar(AA)and clustering while maintaining structure and clustering;sec⁃ondly,the adjacency matrix is mapped to a low⁃dimensional latent space,and the above information is fused using graph regularization to propose three link prediction models:SNMF⁃NEC⁃CN,SNMF⁃NEC⁃AA and SNMF⁃NEC⁃RA;Furthermore,the proposed model parameters are learnt by iteratively updating the rules to obtain the optimal predic⁃tion probability matrix.Comparing with the existing representative methods on six networks,the experimental results show that the AUC and F1 values of the proposed model are improved by 22%and 11.4%,respectively.