Global Coffee & Health: Stress Prediction & Behavioral Clustering

Course: Machine Learning and Data Mining (MLDM)
Academic Year: 2024/2025
Authors: Davide Leone & Filippo Camossi

Project Overview

This project analyzes the Global Coffee Health Dataset (10,000 samples) using a dual-track machine learning architecture to achieve two main objectives:

1. Supervised Modeling (Classification): Build and optimize predictive models to classify an individual's stress level (Low, Medium, High) based on health metrics, demographics, and coffee consumption.
2. Unsupervised Analysis (Clustering): Identify natural behavioral profiles (archetypes) within the population based purely on their habits and characteristics.

The complete end-to-end Data Science pipeline—from Exploratory Data Analysis (EDA) and feature engineering to model evaluation and cluster interpretation—is documented in the main Jupyter Notebook.

Executive Summary & Key Insights

Our analysis uncovered four critical insights regarding the relationship between coffee, health, and stress:

• High Predictive Accuracy: Stress levels can be predicted with high confidence. Our best-performing model, a Bagging Ensemble, achieved a 90.35% accuracy on the test set.
• Prediction Drivers (Health > Coffee): Feature importance analysis for predicting stress revealed that pre-existing health metrics are the dominant factors:
  • Health_Issues: 66% - 91% importance.
  • Health_Risk_Score: ~16% importance.
  • Caffeine/Coffee features: < 5% importance.
• Clustering Drivers (Coffee Defines Groups): Unsupervised analysis (K-Means, k=3) revealed 3 distinct population profiles. The feature importance for creating these groups showed the opposite picture:
  • Caffeine_per_Serving: 53.2% importance.
  • Caffeine_mg & Coffee_Intake: ~28% combined importance.
• The Hidden Relationship: Caffeine consumption is not a good direct predictor of stress. However, it defines behavioral profiles (Cluster 0: High Consumers, Cluster 1: Moderate, Cluster 2: Non-Consumers) which inherently have significantly different stress rates. Health status predicts stress, but the caffeine consumption profile acts as a proxy for the risk group one belongs to.

Methodology & Pipeline

The project follows a rigorous, structured Machine Learning pipeline:

1. Exploratory Data Analysis (EDA): Target distribution analysis (handling class imbalance), correlation mapping (Heatmaps), and statistical testing (ANOVA F-test, Chi-Square) for initial feature selection.
2. Targeted Feature Engineering: Creation of high-impact derived variables, such as Caffeine_per_Serving (handling zero-division cases) and Health_Risk_Score (a composite score of 4 risk factors).
3. Robust Preprocessing & Splitting:
   • Categorical encoding (LabelEncoding, One-Hot).
   • Feature scaling (StandardScaler).
   • Stratified 60-20-20 Split (Train, Validation, Test) to ensure robust evaluation across imbalanced classes.
4. Supervised Modeling: Comparison of 8+ algorithms including Decision Trees, Random Forests, AdaBoost, Gradient Boosting, XGBoost, Bagging, SVM (Linear & RBF), and Multi-Layer Perceptrons (MLP). Evaluated via Accuracy, F1-Score, and Confusion Matrices.
5. Advanced Hyperparameter Tuning: Systematic tuning using Grid Search (for SVM) and Bayesian Optimization via Optuna (for Tree-based models and Neural Networks).
6. Unsupervised Clustering: * Segmented the population using K-Means.
   • Optimal k=3 selected via the Elbow Method (WCSS) and Silhouette Analysis.
   • Cluster visualization using 2D PCA and interpretative profiling via scaled deviation heatmaps.

Tech Stack

• Language: Python
• Environment: Jupyter Notebook / JupyterLab
• Data Manipulation & Math: pandas, numpy, scipy
• Machine Learning: scikit-learn, xgboost, optuna (Bayesian Optimization)
• Visualization: matplotlib, seaborn
• Model Serialization: joblib

🚀 Getting Started

Follow these steps to run the analysis locally.

Prerequisites

• Python 3.8+
• Ensure the dataset CSV file is placed in the root directory of the project.

Installation & Execution

1. Clone the repository:

   git clone https://github.com/SickCiQuattro/MLDM-Coffee-Health.git
   cd coffee-health-ml-analysis

2. Create and activate a virtual environment:

   # On Windows
   python -m venv venv
   venv\Scripts\activate

   # On macOS/Linux
   python3 -m venv venv
   source venv/bin/activate

3. Install the required dependencies:

   pip install -r requirements.txt

4. Launch the Jupyter Notebook:

   jupyter notebook

5. Open MLDM_Project.ipynb and run the cells sequentially to reproduce the analysis.