关键词: 网络表示学习, 拓扑结构, 节点属性, 变分自编码器

Abstract: Network representation learning aims to represent network nodes into a low-dimensional vector space, so that some downstream application tasks such as node classification, link prediction, community detection and recom-mendation can be easily performed by using simple machine learning algorithm. The network nodes usually have attribute information, which is related to the structure information to a certain extent. It is helpful to improve the performance of the downstream tasks by incorporating these information into the learning process of the network representation. However, for different application scenarios, structure information and attribute information are not always linearly correlated, and they are highly nonlinear data. In this paper, variational auto-encoder based network representation learning algorithm is proposed, called VANRL. Variational auto-encoder is a kind of deep neural network, which can not only capture the nonlinear similarity of structure and attribute, but also learn about the distribution of data. In addition, for different application tasks, the flexible fusion of structural information and attribute information makes the learned representation of network nodes achieve satisfactory performance. On four networks, including two social networks and two reference networks, the experimental results show that VANRL can achieve relatively significant results in the tasks of node classification and link prediction.

Key words: network representation learning, topology structure, node attribute, variational auto-encoder