Network embedding has recently attracted lots of attention due to its wide applications on graph tasks such as link prediction, network reconstruction, node stabilization, and community stabilization, which aims to learn the low-dimensional representations of nodes with essential features. Most existing network embedding methods mainly focus on static or continuous evolution patterns of microscopic node and link structures in networks, while neglecting the dynamics of macroscopic community structures. In this paper, we propose a Community-aware Dynamic Network Embedding method (short for CDNE) which considers the dynamics of macroscopic community structures. First, we model the problem of dynamic network embedding as a minimization of an overall loss function, which tries to maximally preserve the global node structures, local link structures, and continuous community dynamics. Then, we adopt a stacked deep autoencoder algorithm to solve this minimization problem, obtaining the low-dimensional representations of nodes. Extensive experiments on both synthetic networks and real networks demonstrate the superiority of CDNE over the existing methods on tackling various graph tasks.

由于网络嵌入在诸如链接预测，网络重建，节点稳定和社区稳定之类的图形任务中得到了广泛的应用，网络嵌入最近引起了很多关注，该目标旨在学习具有基本特征的节点的低维表示。现有的大多数网络嵌入方法主要关注网络中微观节点和链接结构的静态或连续演化模式，而忽略了宏观社区结构的动态变化。在本文中，我们提出了一种考虑社区宏观结构动态的社区感知动态网络嵌入方法（CDNE的缩写）。首先，我们将动态网络嵌入问题建模为总体损失函数的最小化，从而最大程度地保留全局节点结构，本地链接结构，和持续的社区动态。然后，我们采用堆叠式深度自动编码器算法来解决此最小化问题，从而获得节点的低维表示。在合成网络和真实网络上进行的大量实验证明，CDNE在解决各种图形任务方面优于现有方法。