摘要： 针对当前链路预测算法无法有效保留网络图高阶结构特征的问题，提出一种生成对抗式分层网络表示学习算法。根据网络图的一阶邻近性和二阶邻近性，递归地对网络图进行边缘折叠和顶点合并，形成逐层规模变小的子网络图，使用Node2vec算法对规模最小的子网络图进行预处理，并将预处理结果输入到生成对抗式网络（EmbedGAN）模型中，学习得到最小子网络图顶点的低维向量表示，将其输入至上一层子网络的EmbedGAN模型中，作为上一层子网络图顶点的低维向量表示。按照该方法进行逐层向上回溯学习，直至学习到原始网络图，从而得到原始网络图顶点的低维向量表示。在多个不同领域的真实网络数据集上进行链路预测，实验结果表明，该算法的准确率与稳定性均优于LP、Katz和LINE算法。

关键词: 链路预测, 网络表示学习, 邻近性, 生成对抗式网络, 分层网络

Abstract: To address the problem that existing link prediction algorithms cannot effectively retain the high-order structural features of network graph,this paper proposes a generative adversarial hierarchical Network Representation Learning(NRL) algorithm.According to the first-order proximity and second-order proximity of the network graph,the method recursively performs edge collapsing and vertex merging on the network graph to form sub-networks whose scale becomes smaller layer by layer.The Node2vec algorithm is used to pre-process the sub-network with the smallest scale,and the result is input into the Embed Generative Adversarial Network(EmbedGAN) model to learn low-dimensional vector representation of the vertices of the subnetwork graph at the previous level.According to this method,learning process is recursively performed back upward layer by layer until the original network graph is learned,and a low-dimensional vector representation of all vertices of the original network graph is obtained.Experimental results of link prediction on real network data sets in different fields show that the accuracy and stability of this algorithm are better than those of LP,Katz and LINE algorithms.

Key words: link prediction, Network Representation Learning(NRL), proximity, Generative Adversarial Network(GAN), hierarchical network

中图分类号: 

• TP18