Long-tailed classification on graphs is ubiquitous yet challenging in many real-world applications. Recently, oversampling approaches have shown promising performance on imbalanced classification tasks. However, most oversampling methods determine the edge based on the principle of similarity or edge generator, this leads to the generation nodes with poor topological diversity and high homogeneity, which seriously affects the performance of the classifier. To bridge this gap, this paper presents an oversampling method based on Graph Topology Assisted Classifier GAN, named GT-ACGAN. Unlike existing methods, we present a latent variable-based graph topology generation method, to learn the topology of the graph. This approach first pre-trained a VGAE model using the original graph and gets the original graph’s latent variables by the VGAE encoder. Then, the random noises are transformed into latent variables of the fake nodes by neural networks. Finally, the VGAE decoder reconstructs these latent variables into a new adjacency matrix. Then we use the latent variable-based generation method as GT-ACGAN’s generator to generate balanced graph data, and GT-ACGAN’s discriminator is trained to distinguish the authenticity of each node its category. This approach helps the model to generate graph data with typologies diversity and reconstruct erroneous data generation. To validate the effectiveness of the GT-ACGAN, extensive experiments were conducted on manual long-tailed graphs from three classic citation network datasets (Cora, Citeseer, and PubMed) for constructing long-tailed graphs and three different classic graph classifiers (GCN, GraphSAGE, and GAT). GT-ACGAN outperforms state-of-the-art oversampling algorithms for long tail node classification tasks.