Environnement de travail pour le stage ML : Level 1 → Industry Level.
Chaque tâche a son propre dossier, indépendant :
taskXX_nom/
├── data/ # datasets (non versionnés, voir .gitignore)
├── notebooks/ # exploration, prototypage
├── src/ # code réutilisable (.py)
└── outputs/ # modèles sauvegardés, figures, résultats
| Semaine | Tâches | Focus |
|---|---|---|
| 1 | Task 1, 2 | Régression, clustering — bases |
| 2 | Task 3, 4, 5 | Classification multi-classe, données déséquilibrées, reco |
| 3 | Task 6, 7 | Audio/CNN, séries temporelles |
| 4 | Task 8, 9, 10 | Détection d'objets, maintenance prédictive, MLOps |
Les 10 tâches ont été complétées, avec au moins un bonus réalisé sur chacune.
python -m venv venv
source venv/bin/activate # Windows : venv\Scripts\activate
pip install -r requirements.txt| # | Tâche | Niveau | Statut | Résultat clé |
|---|---|---|---|---|
| 1 | Student Score Prediction | Level 1 | ✅ | Le tutorat pèse plus que les heures d'étude (R²=0.27) |
| 2 | Customer Segmentation | Level 1 | ✅ | 5 segments clients, dont un fort potentiel inexploité |
| 3 | Forest Cover Classification | Level 2 | ✅ | 88% accuracy — l'altitude domine tout |
| 4 | Loan Approval Prediction | Level 2 | ✅ | 97% accuracy — le score de crédit pèse 86% |
| 5 | Movie Recommendation System | Level 2 | ✅ | Precision@10 = 0.35 (après correction d'un piège de méthode) |
| 6 | Music Genre Classification | Level 3 | ✅ | 87% accuracy — aucune feature ne domine |
| 7 | Sales Forecasting | Level 3 | ✅ | R²=98.8% — la saisonnalité prime sur l'économie |
| 8 | Traffic Sign Recognition | Industry | ✅ | FPS ok (26), précision limitée par la taille du dataset |
| 9 | Predictive Maintenance | Industry | ✅ | Compromis FDR/Recall — le couple est le meilleur indicateur |
| 10 | MLOps Pipeline | Industry | ✅ | API + Docker + CI/CD validés de bout en bout |
Chaque dossier contient un README.md (contexte technique) et un INSIGHT.md (analyse en langage simple, une page).
Working environment for the ML internship: Level 1 → Industry Level.
Each task has its own, independent folder:
taskXX_nom/
├── data/ # datasets (non versionnés, voir .gitignore)
├── notebooks/ # exploration, prototypage
├── src/ # code réutilisable (.py)
└── outputs/ # modèles sauvegardés, figures, résultats
| Week | Tasks | Focus |
|---|---|---|
| 1 | Task 1, 2 | Regression, clustering — basics |
| 2 | Task 3, 4, 5 | Multi-class classification, imbalanced data, recce |
| 3 | Task 6, 7 | Audio/CNN, time series |
| 4 | Task 8, 9, 10 | Object detection, predictive maintenance, MLOps |
All 10 tasks have been completed, with at least one bonus achieved on each.
python -m venv venv
source venv/bin/activate # Windows : venv\Scripts\activate
pip install -r requirements.txt| # | Task | Level | Status | Key result |
|---|---|---|---|---|
| 1 | Student Score Prediction | Level 1 | ✅ | Tutoring weighs more than study hours (R²=0.27) |
| 2 | Customer Segmentation | Level 1 | ✅ | 5 customer segments, including strong untapped potential |
| 3 | Forest Cover Classification | Level 2 | ✅ | 88% accuracy — altitude dominates everything |
| 4 | Loan Approval Prediction | Level 2 | ✅ | 97% accuracy — credit score weighs 86% |
| 5 | Movie Recommendation System | Level 2 | ✅ | Precision@10 = 0.35 (after fixing a method trap) |
| 6 | Music Genre Classification | Level 3 | ✅ | 87% accuracy — no feature dominates |
| 7 | Sales Forecasting | Level 3 | ✅ | R²=98.8% — seasonality takes precedence over the economy |
| 8 | Traffic Sign Recognition | Industry | ✅ | FPS ok (26), precision limited by the size of the dataset |
| 9 | Predictive Maintenance | Industry | ✅ | FDR/Recall trade-off — torque is the best indicator |
| 10 | MLOps Pipeline | Industry | ✅ | End-to-end validated API + Docker + CI/CD |
Each folder contains a README.md (technical background) and an INSIGHT.md (plain language analysis, one page).