🍷 Wine Quality Prediction

---

📌 Project Overview

This project is a Classification model. By segmenting wines into five distinct quality categories from "Low" to "Very High" this model provides actionable insights for inventory pricing and quality control.

---

📊 The Data

• Source: UCI Machine Learning Repository - Wine Quality Dataset
• Composition: 6,497 observations of Red and White wine.
• Features: 11 physicochemical inputs (pH, Alcohol, Citric Acid, etc.).
• Target: Quality category (Engineered from raw scores).

---

🛠️ Technical Workflow

1. Data Cleaning & Feature Engineering

• Strategic Binning: Transformed raw 0-10 scores into 5 business-relevant tiers: Low, Medium Low, Medium High, High, and Very High.
• Multicollinearity Management: Dropped Residual Sugar to prevent redundancy with Density, ensuring a cleaner feature set.
• Label Encoding: Converted categorical tiers into numerical labels for model compatibility.

2. The "Leakage" Resolution

• Critical Discovery: Identified initial 100% accuracy as Data Leakage (the model was "cheating" by seeing the original quality score).
• The Fix: Stripped all target-related data, forcing the model to rely solely on pure physicochemical benchmarks.

3. Advanced Preprocessing

• Normalization: Applied MinMaxScaler and StandardScaler to handle varying feature magnitudes (e.g., Chlorides vs. Total Sulfur Dioxide).
• Class Balancing (SMOTE): Addressed the rarity of "Very High" quality wines (imbalanced classes) by synthetically oversampling minority classes using SMOTE (Synthetic Minority Over-sampling Technique). This expanded the training set from 1,279 to 2,970 samples.

---

🔬 Modeling & Performance

We utilized an Ensemble-first approach to compare how different architectures handled the chemical complexity of each wine type. To ensure the model works on unseen data, we implemented 5-Fold Stratified Cross-Validation, maintaining class ratios and ensuring a consistent performance margin ($\pm 5%$).

🍷 Red Wine Final Results

Model	Result (Accuracy)	Strategic Insight
Random Forest (Tuned)	68.2%	Top Performer: Effectively mapped complex chemical variances.
KNN Classifier	58.5%	Stability: Established a solid, low-variance baseline.
SMOTE Impact	Balanced	Fairness: Improved prediction on rare premium tiers.

🥂 White Wine Final Results

Model	Result (Accuracy)	Strategic Insight
Random Forest (Tuned)	70.4%	Top Performer: High predictability in acidity/sugar balance.
KNN Classifier	57.4%	Reliable: Effective for high-volume automated sorting.
SMOTE Impact	Robust	Depth: Handled massive sample increases without overfitting.

⚖️ Handling Class Imbalance (SMOTE)

A key challenge in this dataset was the heavy concentration of "mid-range" wines (Quality 5 and 6). To address this, we implemented SMOTE (Synthetic Minority Over-sampling Technique) instead of undersampling.

• Why SMOTE?: We chose this to preserve the rich chemical information within the majority classes. Undersampling would have resulted in a significant loss of data.
• The Result: This approach allowed us to effectively boost the F1-score for the minority classes (high-quality and low-quality wines), ensuring the model identifies "Premium" wines rather than just guessing the most frequent class.

---

🚀 Tech Stack

Languages & Libraries

• Python 3.x
• Pandas & NumPy: Data manipulation and matrix operations.
• Scikit-Learn: Core ML library for scaling, splitting, and modeling.
• Imbalanced-Learn (SMOTE): For handling minority class distribution.
• Matplotlib & Seaborn: For feature importance and correlation heatmaps.

Machine Learning Techniques

• Ensemble Methods: Random Forest, AdaBoost, Gradient Boosting, Bagging.
• Clustering/Proximity: K-Nearest Neighbors (KNN).
• Optimization: Hyperparameter tuning (max_depth, n_estimators), Stratified K-Fold, and SMOTE.

---

The presentation is available here.

Trello dashboard is available here.