基于角色的网络表征学习方法

Role-Based Network Representation Learning Method

网络表征学习技术被广泛应用于获取网络中节点的特征及其语义。已有网络表征学习方法主要研究邻接矩阵或邻接矩阵的幂,使得向量空间中一个节点的相似节点存在于网络中与它相近的局部区域,而未考虑全局区域的结构等价性。根据角色信息,提出基于角色的矩阵分解(Role-MF)模型来获取节点表示。Role-MF模型将角色信息融合在随机游走方法中,在考虑局部信息的同时利用角色信息设计明确的目标矩阵,并通过奇异值分解得到节点表征。实验结果表明,与现有的DWMF、DeepWalk等模型相比,Role-MF模型可以保留结构等价性,当训练比例为10%和90%时,F1值和AUC等各项指标在节点分类和链路预测中都取得了更好的效果。

Network representation learning is widely used to obtain the characteristics and semantics of network nodes.The existing network representation learning methods mainly study the adjacency matrix or the power of the adjacency matrix,making a node in the vector space have similar nodes in the local area approximate to it in the network,but they usually ignore the structural equivalence of the global area.According to role information,this paper proposes a model called Role-Based Matrix Factorization(Role-MF)to obtain node representation.Role-MF integrates role information into a random walk method,uses role information to design a clear target matrix with local information in consideration,and obtains node representation through singular value decomposition.The experimental results show that compared with DWMF,DeepWalk and other existing models,Role-MF can retain structural equivalence,and achieves a higher F1 score and AUC in node classification and link prediction tasks when the training ratio is 10%and 90%.