<p>Real-world classification tasks frequently face highly imbalanced data, where certain classes are represented by disproportionately few instances. This imbalance leads to models favouring the majority class in classification. Consequently, the minority class frequently experiences unsatisfied prediction performance. In multi-class environments, these problems are more noticeable. As the number of classes increases, it becomes more likely to encounter extremely imbalanced situations, where some classes may have very few instances compared to others. One of the techniques to address imbalanced data is via synthetic over-sampling process. However, it is challenging to increase minority class representation without causing class overlap and over-generalisation in the over-sampling process. To address the challenge of extreme class imbalance in multi-class data, we propose the MCDO-BR method (Multi-class Cluster-based Diversity Over-sampling with Boundary Refinement). This method first uses an iterative K-Nearest Neighbours approach to better define the boundaries between classes. It then applies diversity optimisation to generate a wide range of synthetic instances. This helps reducing noise, avoid class overlap, and better handle extremely imbalanced situations by creating more varied and informative samples. MCDO-BR is evaluated on synthetic and real-world datasets and shows better performance than alternative methods.</p>