Network embedding, also known as graph embedding and network representation learning, is an effective method for representing graphs or network data in a low-dimensional space. Most existing methods focus on preserving network topology and minimizing the reconstruction errors to learn a low-dimensional embedding vector representation of the network. In addition, some researchers are devoted to the embedding learning of attribute networks. These researchers usually study the two matrices of network structure and network attributes separately, and then merge them to realize the embedding learning representation of attribute networks. These studies have different performances on a variety of downstream tasks. However, most of these methods have two problems: first, these methods mostly use shallow model to learn structure or attribute embedding, which do not make full use of the rich information contained in the network, such as the neighborhood information of nodes; second, the distribution of the learned network low-dimensional vector representation is overlooked, which leads to poor generalization ability of the model in some real-world network data. Therefore, this paper proposes an adversarially regularized network representation learning model based on attention mechanism, which encodes the topology features and content information of the network into a low-dimensional embedding vector representation through a graph attention autoencoder. Meanwhile, through an adversarial training schema, the learned low-dimensional vector representation could circumvent the requirement of an explicit prior distribution, and thus obtain better generalization ability. Extensive experiments on tasks of link prediction and node clustering demonstrate the effectiveness of learned network embeddings.

网络嵌入，也称为图嵌入和网络表示学习，是一种在低维空间中表示图或网络数据的有效方法。大多数现有方法侧重于保持网络拓扑结构和最小化重构误差以学习网络的低维嵌入向量表示。此外，一些研究人员致力于属性网络的嵌入学习。这些研究者通常将网络结构和网络属性这两个矩阵分别进行研究，然后将它们合并起来，实现属性网络的嵌入学习表示。这些研究在各种下游任务上有不同的表现。然而，这些方法大多存在两个问题：第一，这些方法大多使用浅层模型来学习结构或属性嵌入，没有充分利用网络中包含的丰富信息，如节点的邻域信息；其次，忽略了学习到的网络低维向量表示的分布，导致模型在一些现实世界网络数据中的泛化能力较差。因此，本文提出了一种基于注意力机制的对抗性正则化网络表示学习模型，通过图注意力自动编码器将网络的拓扑特征和内容信息编码为低维嵌入向量表示。同时，通过对抗训练模式，学习到的低维向量表示可以规避显式先验分布的要求，从而获得更好的泛化能力。