Predicting Alumni Engagement Using Institutional Educational Data: A Machine Learning Approach

Abstract

Alumni engagement is a strategic asset for higher education institutions because it supports mentorship, institutional reputation, graduate networking, and long-term sustainability. Despite this importance, many alumni management systems still function mainly as descriptive record systems rather than predictive systems that help institutions identify which graduates are more likely to remain actively connected. This study proposes a machine learning-based Alumni Engagement Prediction System (AEPS) that uses institutional educational data and participation history to predict alumni engagement status. Using institutional alumni records from Paragon International Scool Cambodia, the analytical dataset comprised 474 alumni records and 676 columns, of which 655 were event-participation fields used to derive attendance behavior. The study first transforms raw event attendance into a behavioral participation count and then constructs a binary engagement target using a median split at 12 attended activities. To avoid target leakage, the attendance count is used only to define the target variable and is excluded from the predictor set during model training. The final modeling pipeline applies missing-value imputation, one-hot encoding, stratified train–test splitting, five-fold stratified cross-validation, and hyperparameter tuning across Logistic Regression, Random Forest, Support Vector Machine (SVM), and XGBoost models. Baseline evaluation shows that Logistic Regression achieved the highest cross-validated ROC–AUC among untuned models at 0.7517, while the tuned XGBoost model delivered the strongest held-out test performance with 0.6737 accuracy, 0.6804 F1-score, and 0.7473 ROC–AUC. SHAP-based interpretability analysis further indicates that faculty, university background, generation, and country of studies are the most influential predictors of engagement, while local-versus-abroad study location contributes comparatively little. The findings show that institutional alumni records contain meaningful predictive signals and that machine learning can support more targeted, data-driven alumni relationship management.