Machine learning models can effectively extract meaningful patterns from complex data, but their performance often degrades when the data distribution is highly imbalanced. Class imbalance, where one class contains significantly fewer instances than others, can lead to biased predictions and poor minority-class recognition. To address this issue, this study proposes a hybrid framework combining Area-based Representative Points Oversampling with Shifting (AROSS) and Whale Optimisation Algorithm-SMOTE (WOA-SMOTE). AROSS generates synthetic samples from safe and semi-safe regions, reducing overfitting, while WOA-SMOTE optimises neighbour selection to improve sample diversity and representativeness. The proposed method was evaluated on five imbalanced datasets from the KEEL repository using F1-score, Precision, Recall, and MCC. Experimental results demonstrate that AROSS-WOA-SMOTE achieves competitive performance and outperforms SMOTE, AROSS, Borderline-SMOTE, Safe-Level SMOTE, and WOA-SMOTE on most datasets and evaluation metrics.