This project aims to analyze the chemical composition of wines and identify natural groupings using unsupervised clustering techniques.
Through clustering, we explore whether wines can be categorized without predefined labels, providing insights into similarities and differences in their composition.
✅ Identify natural clusters of wines based on their chemical properties.
✅ Compare multiple clustering algorithms to select the best-performing model.
✅ Evaluate if traditional wine categories emerge naturally from the clustering process.
✅ Develop and deploy a Flask-based web application that allows users to classify wines using the trained model.
- Source: The dataset consists of chemical properties of wines.
- Features: The dataset includes 13 chemical components, such as:
- Alcohol
- Malic Acid
- Ash
- Magnesium
- Flavonoids
- Proline
- ... (see full list in
app.py)
- Performed data cleaning and handled missing values.
- Conducted descriptive statistics and visualizations.
- Computed correlation matrices to understand feature relationships.
- Applied Principal Component Analysis (PCA) to reduce dimensionality.
| Algorithm | Advantages | Disadvantages |
|---|---|---|
| K-Means | Fast, interpretable, efficient for well-separated clusters | Sensitive to initialization, assumes spherical clusters |
| DBSCAN | Detects clusters of various shapes | Highly sensitive to hyperparameters (epsilon, MinPts) |
| Hierarchical (CAH) | No need to predefine k clusters |
Computationally expensive for large datasets |
| GMM (Gaussian Mixture Model) | Handles probabilistic distributions | Less effective for dense clusters |
| BIRCH | Efficient on large datasets | Limited to ellipsoidal cluster shapes |
📌 Final Model Selected: K-Means
✅ Forms well-defined, compact clusters
✅ Computationally efficient
✅ Easy to interpret for analysis and predictions
The models were evaluated using key clustering metrics:
- Silhouette Score: Measures cluster separation (higher = better).
- Calinski-Harabasz Score: Evaluates cluster compactness (higher = better).
- Davies-Bouldin Score: Measures cluster dispersion (lower = better).
📈 K-Means performed best, creating three distinct wine clusters.
📌 DBSCAN failed, classifying 100% of points as noise.
The project includes a Flask-based web app where users can input chemical values of a wine and get its predicted cluster.
✅ User-friendly form to input chemical characteristics.
✅ Real-time prediction of wine cluster using trained K-Means model.
✅ Automated preprocessing (scaling, PCA transformation).
✅ Flask-based deployment for easy usage.
/ML_Wine_Clustering
│── app.py # Flask app for predictions
│── kmeans_model.pkl # Pre-trained K-Means model
│── scaler.pkl # Standard scaler for normalization
│── pca.pkl # Pre-trained PCA model for dimensionality reduction
│── templates/
│ ├── form.html # User input form
│ ├── result.html # Displays predicted cluster
│── static/ # CSS, JavaScript, images
│── Projet_ML.ipynb # Jupyter Notebook with full implementation
│── README.md # Project documentation
pip install -r requirements.txtpython app.pyOpen http://127.0.0.1:5000/ in a browser.
🔧 Future Improvements 🚀 Preprocessing Enhancements:
- Noise reduction and non-linear transformations (Box-Cox, Power Transform).
- Hyperparameter tuning through grid search and cross-validation.
⚡ Alternative Algorithms:
- OPTICS, Graph-Based Clustering, Spectral Clustering.
- Hybrid approach combining ICA and PCA.
📌 Scalability:
- Implement distributed clustering (e.g., MiniBatch K-Means).
- Deploy the model as a cloud-based API.
👤 Author Marwane KASSA
📜 License This project is open-source and free to use.
🔍 Developed as part of a machine learning research on wine clustering.