<p>The growing use of data in education requires reliable, fair, and transparent models to predict student performance, especially in skill-based areas, where identifying high-achieving students enables targeted resource allocation. This research is vital because many educational datasets are small and highly imbalanced, which means that standard machine learning models fail to identify minority groups, such as high-performing students. The study fills a critical gap in the existing literature, where methods tend to ignore privacy, lack transparency, and struggle to handle the extreme class imbalance in small-scale educational data. To address this, we propose a reproducible deep learning framework that combines privacy-preserving data augmentation, adaptive modeling, and interpretability. This approach entails using a Conditional Tabular Generative Adversarial Network (CTGAN) to synthesize realistic data for the underrepresented High Performer class, followed by training a heavily regularized Multi-Layer Perceptron (MLP). The method is evaluated on the Student Physical Education Performance dataset using metrics such as accuracy, weighted F1-score, and class-specific recall. The results show that the proposed MLP achieves perfect recall (1.00) for the High Performer class, with an overall accuracy of 0.827 and a weighted F1-score of 0.828. Compared to baseline methods such as Extreme Gradient Boosting (XGBoost), which failed to detect any high performers (0.00 recall) despite achieving higher overall accuracy, our method is superior for this imbalanced classification task. This work adds to educational analytics by offering a robust, data-centric framework that solves not only the problem of extreme class imbalance but also provides transparent, actionable insights through feature attribution analysis. Feature attribution reveals behavioral factors as stronger predictors of success than physical metrics.</p>