Graph representation learning is a very important part of the field of machine learning. It focuses on learning and extracting representations of nodes and edges in graph data to better analyze features and relationships within the graph structure. Graph neural networks (GNNs) are widely used to deal with node classification problems and has achieved state-of-the-art performance. However, most of the existing methods are performed under the condition that the number of node classes is balanced. In fact, in real-life scenarios, the number of samples of some classes will be significantly less than that of other classes. This leads to insufficient information representation of the minority class graph, thereby affecting the classification performance of GNNs. This paper proposes a new framework GraphAdasyn for imbalanced node classification to address this problem. The model inputs a hybrid view to utilize global and local topology information, expends an over-sampling method that is generalized to multiple-class imbalanced problems in the graph. This method enables adaptively synthesize the number of minority class samples based on the density distribution and is flexibly applicable to various datasets. Through numerous experiments results indicate that proposed framework demonstrates superior performance compared to all baseline methods in imbalanced node classification tasks.