Global Coffee & Health: stress prediction and behavioral clustering

Course project for Machine Learning and Data Mining (2024/2025), by Davide Leone and Filippo Camossi.

The project works on the Global Coffee Health Dataset (10,000 people, 16 columns) and asks two questions:

1. Can we classify a person's stress level (Low, Medium, High) from health metrics, demographics and coffee habits?
2. Do people fall into natural behavioral groups, and how does stress differ between them?

Results

Models are trained on a stratified 60/20/20 split (2,000 test samples) and tuned for macro-F1. Scores below are 5-fold cross-validation on train and validation (8,000 samples), run in MLDM_Project_Improved.ipynb.

Model	Test accuracy	Test macro-F1	F1 on High	CV macro-F1
Decision Tree	87.05%	0.672	0.33	0.659
XGBoost	88.65%	0.659	0.23	0.655
SVM (RBF)	86.65%	0.663	0.29	0.646
Bagging	89.65%	0.638	0.13	0.626
Stacking	90.35%	0.603	0.00	0.603
Random Forest	90.10%	0.604	0.01	0.600

The full table, with AdaBoost, Gradient Boosting and the neural network, is in the notebook.

What the analysis found:

• Accuracy and macro-F1 rank the models in opposite order. About 70% of the test set is Low stress and about 10% is High. Stacking, Random Forest and the neural network reach the highest accuracy (up to 90.35%) but almost never predict High, while the tuned Decision Tree and XGBoost have the best macro-F1. The top five models by macro-F1 sit within one standard deviation of each other.
• Health status predicts stress. Health_Issues carries most of the feature importance in the tree-based models (66% to 91% depending on the model), and Health_Risk_Score, an engineered score built from four risk factors, adds up to about 16%. Direct caffeine features stay under 5%. A Random Forest trained without the two health features drops from 0.60 to 0.34 macro-F1, close to chance for three classes.
• Caffeine shapes the clusters. K-Means with k=3, chosen with the elbow method and silhouette analysis, splits people into heavy consumers (50.0% of the data), moderate consumers (44.4%) and near non-consumers (5.7%). Caffeine_per_Serving accounts for 53.2% of the importance in separating them.
• The two views fit together. Caffeine is a weak direct predictor, but the consumption profiles differ in their share of High stress: 12.1% for heavy consumers, 7.3% for moderate ones and 6.4% for non-consumers. The data is observational, so this shows an association and says nothing about cause.

Note

One feature, Health_Issues, drives most predictions. Treat the absolute scores as properties of this dataset and not as a claim about how well stress can be predicted in general.

Pipeline

1. EDA: target distribution and class imbalance, correlation heatmaps, ANOVA and chi-square tests.
2. Feature engineering: Caffeine_per_Serving (zero-division safe) and Health_Risk_Score.
3. Preprocessing: label and one-hot encoding, standard scaling, stratified split.
4. Classification: decision tree, random forest, bagging, AdaBoost, gradient boosting, XGBoost, stacking, SVM (linear, RBF, sigmoid) and a scikit-learn MLP.
5. Tuning: grid search for the tree, ensemble and SVM models, Optuna (Bayesian optimization) for the neural network.
6. Clustering: K-Means, PCA visualization, profiling with scaled deviation heatmaps.

Repository

File	Content
MLDM_Project_Improved.ipynb	Final analysis. Start here.
archive/MLDM_Project_v1.ipynb	First version. Tunes for accuracy, removes outliers with IQR, trains an unweighted SVM.
archive/MLDM_Presentazione.pdf	Course presentation. It reports the v1 results (for example Bagging 90.35%, SVM 87.80%), which differ from the numbers above.
global_coffee_health_dataset.csv	Dataset.

The notebooks and the presentation are written in Italian.

Run it locally

git clone https://github.com/SickCiQuattro/MLDM-Coffee-Health.git
cd MLDM-Coffee-Health
python3 -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt
jupyter lab

Open MLDM_Project_Improved.ipynb and run the cells in order. A full run takes under ten minutes on a laptop and reproduces the stored results (fixed seeds, tested with the pinned versions in requirements.txt).