Abstract: In recent years, the problem of link prediction in the network has gradually arisen. Compared to traditional heuristic models, link prediction methods based on neural networks have gained the favor of researchers because of the advantage of self-learning. This paper proposes a novel link prediction method based on generative adversarial networks, termed WL-GAN (Weisfeiler-Lehman generative adversarial networks). It first uses a subgraph extraction algorithm and a subgraph encoding algorithm to construct a node pair subgraph, in which the node pair is the structural center. Then it obtains the corresponding adjacency matrix for each node pair of known relationships in the network, which is used to train the generative adversarial networks. A discriminator which can determine whether there is a link in the node pair can be obtained. The experimental results show that WL-GAN achieves excellent performance and stability.

Key words: link prediction, graph labeling, generative adversarial networks (GAN)

摘要： 近些年来，网络中链路预测问题逐渐兴起。相比于传统启发性模型，以神经网络为基础的链路预测方法由于其能够自我学习的优点，逐渐获得研究者的青睐。结合生成式对抗网络，一种创新性的链路预测方法WL-GAN（Weisfeiler-Lehman generative adversarial networks）被提出。WL-GAN首先利用子图提取算法与子图编码算法，为网络中的每条已知关系的节点对构造以该节点对为结构中心的节点对子图，并获得相应连接矩阵。随后，利用矩阵数据来训练生成式对抗网络，最终可以获得能够判断子图中心节点对是否存在链路的判别器。实验结果表明，WL-GAN拥有优秀的性能与稳定性。

关键词: 链路预测, 网络图编码, 生成式对抗网络（GAN）