Class imbalance is a one of the most common problem for supervised classification, particularly in high-stakes domains such as medical diagnosis and fraud detection, in which minority class instances are less in number but their existence is critically important. Standard classifiers then develop a bias toward the majority class, performing poorly on the minority class and potentially leading to catastrophic real-world consequences. For this purpose, we present border-aware Heatmap-based Clustering Under-sampling BA-HCU, a hybrid clustering-based under-sampling framework with mutual information–driven feature weighting. The approach combines K-means clustering, Mahalanobis distance, and density-based scoring to adaptively keep the most informative and representative majority-class samples while achieving balanced class distributions. This design not only improves the classification on imbalanced data but also guarantees diversity in the retained majority class. Comprehensive experiments on 16 benchmark datasets demonstrate that BA-HCU outperforms baseline under-sampling methods consistently, with superior ROC AUC, PR AUC, and Partial ROC AUC scores. While ROC AUC provides an overall picture of performance, PR AUC is sensitive to the detection of the minority class, and Partial ROC AUC enables focused analysis on crucial decision thresholds. These measures give a complete evaluation of classifier performance in imbalanced learning situations.