Bayesian Networks Framework For Predicting Academic Success: A Study With Qassim University Freshman Students In STEM Courses

Abstract

Typically, models for predicting students’ academic success achieve high classification accuracy but offer poor explainability for administrators and teachers who must interpret and act on their predictions. To predict students’ academic success and provide interpretable explanations of their predictions, this study proposes and experimentally evaluates a hybrid machine learning framework that combines an ensemble learner, Bayesian Networks (BNs) structure learning, and post-hoc explainable artificial intelligence (XAI) techniques. The academic success was treated as a binary outcome (final grade G3 ≥ 10 versus G3 < 10) using a subsample of 95 secondary school graduates from the publicly available Saudi Student Performance Dataset. These students were inducted into Qassim University, Saudi Arabia in STEM courses. All classifiers were trained with a 75%/25% train/test split and evaluated using 5-fold cross-validation, reporting accuracy, precision, recall, F1-score, and ROC AUC. LightGBM achieved the best held-out performance (accuracy = 0.875, F1 = 0.897, and ROC-AUC = 0.926). In this small sample, prior academic achievement (represented by first-, second-period grades (G1, G2) and success represented by G3 ≥ 10) was the only statistically significant predictor of success (p < 0.001), while lack of school educational support and social engagement outside of school showed suggestive associations but failed to reach the α = 0.05 significance threshold. The SHAP and LIME explanations were similar and identified the same key drivers of the model's predictions: prior GPA, social engagement outside of school, and weekly study time. A Bayesian network learned using the hill-climb search with the Bayesian Information Criterion (BIC) score recovered a single, directed edge from prior GPA to academic success, providing both explanatory power and structural sparsity in probabilistic graphical models from a limited dataset. The results illustrate that such a hybrid approach, which integrates ensemble learning with Bayesian network modeling and post-hoc XAI methods, yields an interpretable and structurally sparse probabilistic model despite the small sample size.