Graph Embedding Method for Node Classification Guided by Structure and Attributes

The node category in the graph usually has a strong correlation with the node attribute and the domain. The embedding vector learned by the random walk method only considers the topology of the node, ignoring the category attribute of the node, which reduces the accuracy of the node classification task. This paper proposes a network representation learning method that integrates node attributes and structural features. In the process of node sequence sampling, the method selects one node as a follow-up from the neighbor nodes whose similarity with the current node is higher than the threshold. And finally submits a sequence of a certain length to the model for training, and gets the embedding vector for each node. The experimental results of node classification on public datasets show that this method has achieved good results on training sets of different sizes. The macro-F1 value of node classification is 5.1% and 3.6% higher than the optimal baseline method, respectively.