Predicting student performance using educational data is a significant area of machine learning research. However, class imbalance in datasets and the challenge of developing interpretable models can hinder accuracy. This study compares different variations of the Synthetic Minority Oversampling Technique (SMOTE) combined with classification algorithms to create prediction models. The results show that SMOTE with Edited Nearest Neighbors is superior, and the balanced random forest classifier performs better when using SMOTE-ENN, achieving 96% accuracy, precision, and F-value. Smote also has faster execution time. For model interpretability, combining Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) provides deeper insights. LIME is suitable for single-prediction interpretation, while SHAP is better for overall model interpretation. This research offers guidelines to mitigate data imbalance and improve fairness in education through data-driven innovations like early warning systems. It also educates academics on explainability approaches to facilitate wider use of machine learning methods.