A Multi-layer Perceptron Framework for Pre-admission Disciplinary Risk Prediction in Nigerian Universities: A Fairness-aware Approach Using Synthetic Data

Nigerian universities have long struggled to identify, before matriculation, which applicants are likely to become disciplinary problems on campus. Existing screening procedures are largely manual and reactive — problems surface after enrolment rather than at the gate. To close this gap, we built and tested a Multi-Layer Perceptron classifier that assigns each applicant a probabilistic risk score at the point of admission, giving institutional officers an evidence-based basis for early counselling and targeted resource deployment, well before misconduct occurs. Working with real student records was not feasible. Nigeria's data protection obligations impose strict constraints on profiling, and authentic admission data from universities was unavailable for this research. We therefore generated 5,000 synthetic applicant profiles using a Conditional Tabular GAN, a method purpose-built for datasets that mix continuous, ordinal, and categorical variables. Three statistical tests — the Kolmogorov–Smirnov statistic, Wasserstein distance, and Jensen–Shannon divergence — confirmed that the synthetic profiles reproduced the structural properties of realistic admission populations with high fidelity. Under five-fold stratified cross-validation, the MLP returned an accuracy of 0.841 ± 0.018, F1-score of 0.812 ± 0.021, and AUC-ROC of 0.891 ± 0.014, outperforming Logistic Regression, Random Forest, SVM, and XGBoost across all reported metrics. Two findings deserve particular attention. First, SHAP attribution analysis singled out prior disciplinary record and JAMB score as the variables driving predictions most strongly — a result with direct implications for what admissions officers should scrutinize. Second, the model treated applicants from different geopolitical zones unequally; an EOD of 0.078 across zones exceeded acceptable thresholds. Fairness-regularized retraining brought that Figure down to 0.043 with less than one percentage point of accuracy lost. To prevent the system from operating as a black box, a three-tier human-in-the-loop review protocol is proposed, and the entire deployment framework is mapped against Nigeria's National Data Protection Regulation and Data Protection Impact Assessment requirements.