Class imbalance severely degrades a classifier’s ability to recognize minority-class instances, whose correct identification is often the central goal in high-stakes domains such as fraud detection and medical diagnosis. Classic data-level remedies such as SMOTE and its adaptive variant ADASYN improve performance by synthesizing additional minority examples, yet they still rely on simple heuristics to decide where and how many samples to generate, leading to noisy or poorly placed data in complex manifolds. We introduce GAT-ADASYN, a graph-attention-guided oversampling method that replaces these heuristics with a lightweight, fixed Graph Attention Network (GAT) module. Attention weights computed between each minority sample and its K nearest neighbours serve (i) to quantify learning difficulty via majority-weighted scores that drive the per-sample synthesis quota, and (ii) to form a probability distribution that steers interpolation toward semantically relevant neighbours—regardless of class—thus populating sparse boundary regions without excessive noise. Extensive experiments on five public imbalanced datasets show that, when paired with a Random-Forest classifier, GAT-ADASYN achieves the highest mean F1-Score, G-mean and AUC across all baselines. Additional visualizations on synthetic Moon and Circle datasets reveal that our method produces well-situated, manifold-respecting synthetic clusters. The results demonstrate that embedding graph attention into the oversampling pipeline is an effective and computationally inexpensive strategy for mitigating class imbalance.