Unpacking the Black Box: A Hybrid XAI Framework for AutoGluon-Based Multiclass Student Outcome Prediction

High student dropout rates remain a significant impediment to achieving the United Nations SDG 4 (equitable education). While Artificial Intelligence (AI) offers robust early risk prediction, the intrinsic black-box nature of high-performing models constrains their transparency. This study designs and investigates a multi-layered Explainable AI (XAI)-based assessment framework to generate actionable insights for student retention. We utilized AutoGluon to construct high-performing multiclass classification models (Graduated, Dropout, or Enrolled) on a higher education dataset. To address the complexity of the AutoGluon-generated models, we employed a hybrid XAI framework that couples global interpretability via a decision tree surrogate model and local interpretability via LIME (Local Interpretable Model-agnostic Explanations). The analysis revealed that models from the Boosting family, particularly XGBoost with bagging level 2, achieved the highest predictive performance (exceeding 0.890 across all metrics). The global analysis demonstrated that academic factors were the primary drivers of prediction, but critical socio-economic factors, such as Tuition fees, also exerted significant influence. Local LIME analysis provided granular, case-specific insights, strongly linking dropout status to first-year academic challenges and to features such as age at enrollment. This integrated XAI approach transforms complex models into an interpretable system, supporting student retention and educational equity (SDG 4).