Predict. Explain. Intervene. An end-to-end AI platform that identifies at-risk university students before it's too late.
UNI-AI is a full-stack machine learning platform built on the OULAD dataset (32,593 students, 28 engineered features across 7 tables). It trains and compares 9 classifiers, tunes the top models with Optuna, builds a soft-voting Ensemble, and explains every prediction using SHAP — all wrapped in a polished interactive Streamlit dashboard.
git clone https://github.com/hagerbayoumi11/UNI-AI.git
cd UNI-AI
pip install -r requirements.txt
streamlit run app.pyDownload OULAD CSV files from analyse.kmi.open.ac.uk/open-dataset and place them in the root directory.
Real-time overview of all 32,593 students — pass rates, withdrawal rates, risk distribution, engagement heatmap, top at-risk students, model leaderboard, and AI-generated insights.
Enter any student profile across 4 dimensions (Demographics, VLE Engagement, Assessment Performance, Registration Timing) and get an instant dropout probability with SHAP feature contributions and personalized recommendations.
Full comparison of 9 classifiers + Optuna-tuned variants + Soft-voting Ensemble — visualized as bar charts, ROC curves, and radar plots.
Exploratory analysis across all 7 OULAD tables — result distribution, pass rate by education level, gender breakdown, IMD band impact, VLE engagement patterns, and feature importance.
Global beeswarm plots, single-student waterfall charts, and dependence plots — showing exactly which features drive each prediction.
| Rank | Model | Accuracy | F1 Weighted | ROC-AUC | CV F1 |
|---|---|---|---|---|---|
| 1 | XGBoost (Tuned) | 92.01% | 92.01% | 97.92% | 92.16% |
| 2 | LightGBM (Tuned) | 91.61% | 91.61% | 97.87% | 91.80% |
| 3 | Ensemble (Top 3) | — | 88.80% | 93.10% | — |
| 4 | Gradient Boosting | 91.41% | 91.41% | 97.78% | 91.59% |
| 5 | XGBoost | 91.09% | 91.09% | 97.73% | 91.28% |
| 6 | Random Forest | 91.00% | 91.00% | 97.53% | 91.15% |
| 7 | SVM | 90.98% | 90.99% | 96.84% | 91.04% |
| 8 | Decision Tree | 90.96% | 90.97% | 97.02% | 91.33% |
| 9 | LightGBM | 90.75% | 90.76% | 97.64% | 90.91% |
| 10 | KNN | 89.88% | 89.88% | 96.21% | 89.49% |
| 11 | Logistic Regression | 89.78% | 89.79% | 95.89% | 89.90% |
| 12 | MLP Neural Net | 88.56% | 88.56% | 96.45% | 89.24% |
Tuning: Optuna TPE — 25 trials LightGBM + 25 trials XGBoost
| Layer | Tools |
|---|---|
| Data Processing | Python, Pandas, NumPy |
| Feature Engineering | 28 features across 7 OULAD tables |
| ML Models | XGBoost, LightGBM, Random Forest, Gradient Boosting, SVM, KNN, MLP Neural Net, Logistic Regression, Decision Tree |
| Hyperparameter Tuning | Optuna TPE (50 trials) |
| Ensemble | Soft-voting Top-3 + optimal threshold |
| Explainability | SHAP (Beeswarm, Waterfall, Dependence) |
| Frontend | Streamlit multi-page app |
| Visualization | Plotly, Matplotlib, Seaborn |
UNI-AI/
├── app.py # Main Streamlit entry point
├── project.ipynb # Full ML pipeline & experiments
└── pages_code/
├── dashboard.py # Overview dashboard
├── predict.py # Student risk prediction
├── model_perf.py # Model leaderboard & comparison
├── eda.py # Exploratory data analysis
├── shap_page.py # SHAP explainability
└── utils.py # Shared utilities
OULAD — Open University Learning Analytics Dataset
- 32,593 students | 28 engineered features | 7 relational tables
- Download: analyse.kmi.open.ac.uk/open-dataset
Hager Bayoumi — Data Scientist & ML Engineer
