🥗 NutriScan – Predicting Nutrition Grades of Food Products COE305 – Machine Learning Final Project Team Members: Eda Şahin Tuana Harmankaya Zehra Özcan
📌 Project Overview NutriScan is a multi-class machine learning system designed to predict the nutrition_grade_fr (A–E) of food products using structured nutritional values per 100g. Food databases often contain missing or inconsistent nutrition grades. This project builds a robust predictive pipeline to estimate grades automatically using numeric nutrient features. Final best-performing model: Random Forest
🎯 Problem Statement Given nutritional values per 100g: energy_100g fat_100g saturated-fat_100g sugars_100g salt_100g proteins_100g fiber_100g carbohydrates_100g Predict: nutrition_grade_fr ∈ {a, b, c, d, e} This is treated as a multi-class classification problem.
📊 Dataset Original Source: OpenFoodFacts (Kaggle) Customized subset: 5,000 US products Final labeled samples: 4,026 Data type: Structured (tabular) Important: The column nutrition-score-fr_100g was excluded to prevent data leakage.
⚙️ Data Preprocessing Removed rows with missing target labels Median imputation for numeric missing values StandardScaler applied for linear/distance-based models Stratified 80/20 train-test split Stratified K-Fold (3-fold) used in hyperparameter tuning Evaluation metrics: Accuracy Macro Precision Macro Recall Macro F1 (primary metric due to class imbalance)
🤖 Models Implemented Non-Ensemble Models Logistic Regression Decision Tree K-Nearest Neighbors (KNN) Ensemble Models Random Forest Gradient Boosting Tuned Random Forest (GridSearchCV)
📈 Model Performance (Test Set) Model Accuracy Macro F1 Logistic Regression 0.629 0.606 Decision Tree 0.819 0.821 KNN 0.667 0.659 Random Forest 0.857 0.854 Gradient Boosting 0.851 0.847 Tuned Random Forest 0.857 0.854
Best Model: Random Forest Tree-based ensemble models significantly outperformed linear baselines due to their ability to capture non-linear relationships and feature interactions.
🔍 Feature Importance Insights Model-based feature importance analysis showed: salt_100g saturated-fat_100g energy_100g sugars_100g These were the most influential predictors. This aligns with nutrition scoring logic: higher salt and saturated fat levels generally correspond to worse grades.
📊 SHAP Interpretability SHAP (TreeExplainer) was used to analyze feature contributions. Findings: Strong interactions between salt, saturated fat, and energy Tree-based models effectively captured non-linear feature interactions
📉 Confusion Matrix Insights Most classification errors occurred between neighboring grades (e.g., b↔c, c↔d). Extreme grades (a vs e) were rarely confused, indicating strong separation of very healthy vs very unhealthy products.
💻 User Interface (Streamlit) A Streamlit web application was implemented to: Browse 4,026 products Filter by grade Search by product name Display nutritional values Show health risk score (mapped from grade) Provide rule-based nutritional insights Run UI Locally pip install -r requirements.txt streamlit run app.py Then open: http://localhost:8501
🛠 Technologies Used Python pandas numpy scikit-learn matplotlib seaborn shap streamlit Google Colab
🚀 Future Improvements Incorporate ingredient text using NLP Apply class imbalance techniques Probability calibration Explore advanced boosting libraries (XGBoost, LightGBM)
📂 Repository Structure (Suggested) ├── notebook.ipynb ├── app.py ├── requirements.txt ├── README.md └── data/
📚 References OpenFoodFacts Dataset Breiman (2001) – Random Forest Friedman (2001) – Gradient Boosting Scikit-learn Documentation