Abstract:
Machine learning predictive approach is created to evaluate the impact of varying factors on school desertion in Narok West Sub-County institutions through analysis of local population dynamics.The lack of evidence-based evaluation in dropout prevention programs within Narok West Sub County hampers the effectiveness of bursaries and school feeding programs together with community awareness measures although they continue to persist as significant issues.Current interventions are unsuccessful because they work on a cure-after-dropout method instead of making forecasts to stop it beforehand. This paper constructs and proves a machine-learned system to measure the weight of exposures to major mitigation variables on the risk of dropping out of primary school in Narok West Sub-County, Kenya. Structured questionnaires were used to gather field data (at the start (n=399) and scaled to 1,000), which was supplemented by Monte Carlo simulation; following cleaning and preprocessing, the analytic dataset included 9,796 records containing 16 variables (demographics, access, and interventions (bursaries, school feeding, community sensitization, and economic empowerment). An 80/20 split (with 5-fold cross-validation) was trained on three supervised models; Random Forest (RF), Support Vector Machine (SVM), and XGBoost and assessed with Accuracy, Precision, Recall, F1-score and ROC-AUC. XGBoost had the highest discrimination (AUC 0.804; F1 0.771; Accuracy 74.0%), then RF (AUC 0.790; Accuracy 72.9%), and then SVM (AUC 0.751). Model-agnostic Permutation Importance and SHAP (Shapley Additive Explanations) were used in order to provide transparency and interpretability of the policies. Across methods, bursary receipt and bursary amount continued to be the strongest protective mitigators, with increased distance to school and increased contributions towards monthly fees contributing to risk of dropout; school feeding participation and meals per day and attendance at education-oriented community meetings had other more minor protective effects. The proven model allows ranking the interventions based on evidence and allocating resources accordingly.
