Academic Risk Profiling and Tailored Student Guidance Through Machine Learning in Kinshasa Universities

Abstract

Educational data can help identify, before the end of a semester, academic trajectories that deserve timely attention. This study develops an academic risk profiling and tailored student guidance using machine learning techniques for universities in Kinshasa. A quantitative, experimental, and predictive design was applied to a harmonized dataset of 9,000 student-semester observations and fourteen predictors after removing a redundant composite engagement index. Data preparation, modelling, and visualisation were performed in a reproducible Python/Jupyter Notebook workflow using pandas, NumPy, scikit-learn, and Matplotlib. A stratified 80/20 training-test split and five-fold stratified cross-validation were used to compare multinomial logistic regression, decision tree, random forest, and multilayer perceptron models. On the independent test set, the multilayer perceptron achieved the highest accuracy (0.720) and macro-AUC-ROC (0.945), while logistic regression achieved the highest balanced accuracy (0.714). Random forest was retained because it achieved the highest macro F1 score (0.686), the prespecified criterion for protecting attention to minority classes, together with macro precision of 0.713 and macro One-vs-Rest AUC-ROC of 0.941. Continuous assessment average, midterm average, and the previous validated-credit rate were the most informative signals. Recommendations are derived from the predicted class, while uncertain cases are flagged for human review. The proposed model is therefore a decision-support tool rather than an automated decision-maker.