Predicting Student Dropout Risk Using Synthetic Data
2024-12-31
Project Overview
🖥️ Live App: View Student Dropout Predictor Dashboard
Objective: Build a tool to predict university student dropout risk after their second semester.
Built with: - R, ggplot2, plotly, Shiny - ML models: Random Forest, Logistic Regression, GBM - SMOTE for class imbalance
Data: Synthetic, designed to replicate academic and behavioral student signals.
Spring 2025 Proposal
I Promised To: - ✅ Build a predictive AI based on attendance and Fitbit activity - ✅ Address cybersecurity and ethics in education - ✅ Deliver a working demo with interpretable features - ✅ Reflect on spiritual implications of AI (briefly)
The Dataset
- 150 synthetic students
- Variables include:
- Course_Count, GPA, EFC
- Fitbit-like data: Fatburn, Cardio, Cardio Peak
- Dropout defined as no enrollment after 2nd semester
Preprocessing: Wide format + engineered changes.
Feature Engineering
- GPA_Change = GPA2 - GPA1
- Attendance_Change = Course_Count2 - Course_Count1
- EFC_Change, Cardio_Change, etc.
Used to capture student momentum or decline over time.
Modeling Strategy
- Models Tested:
- ✅ Random Forest
- Logistic Regression (baseline)
- Gradient Boosting (best AUC)
- Imbalance Handling:
- SMOTE for minority class boost
- 5-fold repeated CV for robust validation
Model Performance Summary
| Model | Accuracy | AUC | Kappa |
|---|---|---|---|
| Random Forest | 36.7% | 0.60 | -0.28 |
| Logistic Regression | 53.3% | 0.54 | ~0.03 |
| Gradient Boosting | 46.7% | 0.67 | -0.09 |
Insight: GBM had the best discrimination power; Logistic Regression worked as expected for a baseline.
📊 Visualizing Risk Levels
Interpreting the Risk Plot
- Bars show predicted dropout probability for top 20 students
- Color legend:
- 🟢 True Positive: correctly predicted dropout
- 🔴 False Negative: dropout missed by model
- 🔵 True Negative: correctly predicted to stay
- 🟠 False Positive: flagged dropout, but stayed
- Useful for advisors to review cases & intervene early
Ethics & Cybersecurity
- All data pseudonymized
- Protected processing and secure data handling
- Ethical reflections:
- AI supports human care, not replaces it
- Christian ethics: fairness, transparency, stewardship
“AI should serve the student, not replace the counselor.”
Future Work 🚀
- Predict dropout during first semester
- Requires daily logs: LMS logins, activity, device data
- Add real-time flags in the dashboard
- Train models on institutional data for higher fidelity
Thank You!
Let’s build systems that notice students before they disappear.
🧠 Questions? 💬 Contact: [vfashina@oru.edu]