关键词: 关键词: 网络嵌入, 表示学习, 随机游走, 网络采样, 属性网络, 节点分类

Abstract: Abstract: It is an important task to carry out inference calculation on graph structure data. The main challenge of this task is how to represent network knowledge so that machines can easily understand and use graph structure data. After comparing with the existing representation learning models, it is found that the model based on random walk methods are likely to ignore the special effect of attributes on the adjacency relationship between nodes. There for, a hybrid random walk method based on node adjacency and attributes association was proposed. The basic idea is to calculate the attribute weight through the common attribute distribution among adjacent nodes, and obtain the sampling probability from node to each attribute. Then the network information can be extracted from adjacent nodes and non-adjacent nodes with common attributes. Finally, the network representation learning model based on node attribute bipartite graph was constructed to learn nodes’ embeddings through the above sampling sequence. Experimental results on Flickr, BlogCatalog and Cora show that the average accuracy of node classification by the model is 89.07%，which was 2.13 percentage points higher than recent work and 21.34 percentage points higher than classical work. By comparing different random walk methods, it is also found that increasing the sampling probability of attributes that promote node association can improve the information contained in the sampling sequence.

Key words: Keywords: network embedding, representation learning, random walk, network sampling, attributed network, node classification