Machine learning models for academic performance prediction: interpretability and application in educational decision-making

The integration of artificial intelligence in education has enabled the development of predictive models for academic performance. However, most existing approaches lack interpretability and do not provide actionable insights for decision-making. This study addresses these limitations by deploying optimized machine learning models, specifically XGBoost and Random Forest, to predict student performance considering geographically, institutional, socioeconomic, and academic factors. Unlike previous research focused only on accuracy, this work incorporates SHAP-based interpretability techniques and an interactive decision support system to analyze the impact of various variables on educational outcomes. The model was trained and validated on a dataset of 50,000 student records, optimized through hyperparameter tuning and cross-validation. Results indicate that XGBoost achieves an R 2 of 0.91, outperforming traditional approaches, and reduces the mean square error (MSE) by 15%. The feature importance analysis reveals that five variables explain 72% of the variability in performance, highlighting the influence of socioeconomic conditions, infrastructure, and the student-teacher ratio. In addition, simulations of educational policies show that improving teacher training and access to technology increases performance by 18% and reduces dropout by 12%. This study presents a scalable and interpretable predictive model that anticipates student performance and helps optimize educational strategies through artificial intelligence applied to decision-making.