融合高阶路径和非负矩阵分解的链路预测

Link Prediction Combining High-Order Path and Non-Negative Matrix Factorization

现有的大多数加权网络链路预测方法仅考虑了自然权重,忽略了高阶加权路径信息贡献,导致预测准确度降低。为此,提出了融合高阶路径和非负矩阵分解的链路预测模型。首先,将任意权重网络的邻接矩阵映射到低秩隐特征空间,以捕获一阶权重链接信息;其次,根据节点强度计算链接权重强度,以获得局部相似度矩阵,并将其扩展为高阶路径相似度;最后,启用图正化方法耦合二阶和高阶路径相似度,以保持局部和全局结构。此外,利用拉格朗日乘法规则学习模型参数来获得最优因子矩阵。在6个典型的实际加权网络上与现有的主流加权网络链路预测方法进行比较分析,得到本模型性能最优。

Most of the existing weighted network link prediction methods only consider the natural weight and ignore the contribution of high-order weighted path information,which leads to the degradation of the prediction accuracy.To address this problem,a link prediction model that fuses natural weight links and higher-order path information with non-negative matrix decomposition is proposed.Firstly,the model maps the adjacency matrix of an arbitrary weight network to a low-rank hidden feature space to capture first-order weight link information.Secondly,the local similarity matrix is obtained by calculating the link weight strength based on the node power strength,and then the local similarity matrix is extended by the higher-order path similarity.Finally,the graph normalization method is enabled to couple the second-order and higher-order path similarities to maintain the local and global structures.In addition,Lagrange multiplication rules are used to learn the model parameters to obtain the optimal factor matrices.The experimental results on six typical real weighted networks compared with the existing mainstream weighted network link prediction methods show that the proposed model has the optimal performance.