<p>In recent years, Graph Neural Networks (GNNs) have demonstrated exceptional performance in graph node classification tasks by effectively modeling complex relationships between nodes. However, real-world graphs often exhibit significant class imbalance, with minority class nodes being underrepresented, which leads to biased classification results that favor majority class. To address this challenge, we propose GraphCKSA, a novel graph resampling framework designed to tackle imbalanced node classification by CENN-KCQ resampling and dual-view edge optimization strategies. Specifically, our contributions are as follows: (i) We propose an innovative CENN undersampling method to address issues of noise and over-compression in majority class nodes, improving undersampling accuracy and efficiency. (ii) We design a KCQ-SMOTE oversampling method that combines KMeans clustering with intra-cluster SMOTE, while leveraging Q-learning to intelligently determine the optimal number of clusters K. This ensures high-quality and diverse oversampling of minority class nodes. (iii) We propose a dual-view edge optimization strategy, optimizing edge connections from both local and global perspectives. Locally, a resampled graph structure is built using cluster centers and k-nearest neighbors. Globally, a graph attention network reinforces key edges and filters out irrelevant ones. GraphCKSA constructs a high-quality, balanced augmented dataset, which significantly boosting the performance of GNN in imbalanced node classification tasks. Experimental results on three public benchmarks—Cora, Citeseer, and PubMed—demonstrate that GraphCKSA outperforms state-of-the-art models across ACC, AUC-ROC, and F1. Further ablation studies, imbalance ratios, oversampling scales and hyperparameter analysis validate the robustness and effectiveness of GraphCKSA, showcasing its potential for addressing imbalanced graph classification challenges.</p>