Few-shot learning has been solving classification problems with few labels in the image and text domains, which has achieved great success in such Euclidean domains. However, a large amount of data can be represented as attributed networks defined by non-Euclidean domains, and fewer studies have successfully applied few-shot learning to attributed networks. Firstly, they ignore the relationship between nodes, so they can not obtain an adequate and accurate representation of nodes. Secondly, the use of an artificial set distance function to measure the similarity between nodes is defective. Because some special samples can cause perturbation to the classifier. In order to solve the above problems, we propose a new framework based on relation networks - graph relation networks (GRN). Specifically, GRN first sets up an episode training mechanism to simulate a real-world few-shot scenario, then uses graph attention networks to obtain a representation of each node with importance scores, and finally learns to compare the similarity of unclassified and classified nodes through the relation network. We conduct extensive experiments on real-world datasets. The experimental results show that the method improves the performance of graph node classification.