Node classification for class-imbalanced graph data presents significant challenges, as existing Graph Neural Networks (GNNs) typically either learn node embedding on class balanced graph data or ignore such class imbalance. This leads to predictive bias in terms of favoring majority classes while under-representing minority classes. To address this issue, this study proposes an innovative algorithm named GNN-GAN, which integrates GNN with generative adversarial network (GAN), where the GNN extracts latent features from input node attributes that are balanced through a conditional GAN and data fusion strategy. The proposed GNN-GAN bridges the gap between node classification and the class imbalance learning, which is nontrivial in the context of often class-imbalanced real-world graph data. Experiments on several classic graph datasets shows that the GNN-GAN out-performs current state-of-the-art baselines and are robust to graph datasets with varying structures and imbalance ratios.