Graph convolutional network (GCN) has made remarkable progress in learning good representations from graph-structured data. The layer-wise propagation rule of conventional GCN is designed in such a way that the feature aggregation at each node depends on the features of the one-hop neighbouring nodes. Adding an attention layer over the GCN can allow the network to provide different importance within various one-hop neighbours. These methods can capture the properties of static network, but is not well suited to capture the temporal patterns in time-varying networks. In this work, we propose a temporal graph attention network (TempGAN), where the aim is to learn representations from continuous-time temporal network by preserving the temporal proximity between nodes of the network. First, we perform a temporal walk over the network to generate a positive pointwise mutual information matrix (PPMI) which denote the temporal correlation between the nodes. Furthermore, we design a TempGAN architecture which uses both adjacency and PPMI information to generate node embeddings from temporal network. Finally, we conduct link prediction experiments by designing a TempGAN autoencoder to evaluate the quality of the embedding generated, and the results are compared with other state-of-the-art methods.

图卷积网络（GCN）在从图结构化数据学习良好表示形式方面取得了显着进步。常规GCN的逐层传播规则的设计方式是，每个节点的特征聚合取决于一跳相邻节点的特征。在GCN上添加关注层可以使网络在各个一跳邻居中提供不同的重要性。这些方法可以捕获静态网络的属性，但不适合捕获时变网络中的时间模式。在这项工作中，我们提出了一个时态图注意力网络（TempGAN），其目的是通过保留网络节点之间的时域接近性来从连续时间时态网络中学习表示形式。第一的，我们在网络上执行时间漫游，以生成表示各个节点之间时间相关性的正点向互信息矩阵（PPMI）。此外，我们设计了一个TempGAN架构，该架构使用邻接和PPMI信息来从时态网络生成节点嵌入。最后，我们通过设计TempGAN自动编码器进行链接预测实验，以评估生成的嵌入的质量，并将结果与​​其他最新方法进行比较。