🍷 White Wine Quality Classification

Data Mining

Author: Nathaly Ingol

Toolkit: Orange 3

---

📌 Project Overview

This project explores classification techniques using the White Wine Quality dataset (4,898 instances) from the UCI Machine Learning Repository.

The dataset contains:

• 11 numerical physicochemical features:

  • Fixed acidity
  • Volatile acidity
  • Citric acid
  • Residual sugar
  • Chlorides
  • Free sulfur dioxide
  • Total sulfur dioxide
  • Density
  • pH
  • Sulphates
  • Alcohol

• Target variable:

  • Wine quality score (3–9)

The analysis was performed using Orange 3, a visual drag-and-drop data mining toolkit built on Python and scikit-learn.

Time spent on this assignment: ~6–7 hours

---

🧹 Dataset Preparation

1. Loaded dataset using the File widget.

2. Used the Edit Domain widget to:

   • Convert quality from Numeric → Categorical
   • Set quality as the Target variable

⚠️ This step was essential. Without converting the target to categorical, Orange treated the problem as regression and produced errors (e.g., Naive Bayes requires a categorical class).

📊 Dataset Characteristics

• 4,898 instances

• No missing values

• Class imbalance present

  • Classes 5 and 6 account for ~74% of the data
  • Classes 3 and 9 are rare (20 and 5 samples)

Because of this imbalance, accuracy alone is not a reliable metric.

---

🌳 Decision Tree Analysis

A Decision Tree learner was connected to a Tree Viewer widget for visualization.

Key Findings:

• The root split was on alcohol content
• Alcohol is the most important predictor of wine quality
• Unlimited tree depth → Overfitting
• Max depth = 5 → Better interpretability and generalization

The Tree Viewer clearly showed the branching logic of the model.

---

📈 Model Comparison (5-Fold Stratified Cross-Validation)

Six models were evaluated using the Test & Score widget.

Model	AUC	Accuracy	F1	Precision	Recall	MCC
Neural Network	0.766	55.3%	0.538	0.540	0.553	0.309
CN2 Rule Induction	0.736	61.1%	0.611	0.610	0.611	0.423
Decision Tree	0.713	57.7%	0.576	0.577	0.577	0.374
Naive Bayes	0.690	44.2%	0.446	0.464	0.442	0.210
SVM	0.689	42.5%	0.413	0.420	0.425	0.161
kNN	0.669	47.5%	0.462	0.460	0.475	0.193

🔎 Interpretation

• Neural Network

  • Highest AUC (0.766)
  • Best overall class discrimination
  • Statistically significantly better than other models

• CN2 Rule Induction

  • Highest raw accuracy (61.1%)
  • Produces human-readable IF–THEN rules
  • Strong balance of interpretability and performance

• Decision Tree

  • Good performance (57.7%)
  • Fully interpretable

• SVM & kNN

  • Underperformed
  • Likely due to lack of feature normalization
  • Distance-based models are sensitive to feature scale

  

---

🔬 Unsupervised Learning (Bonus)

k-Means Clustering (k = 3)

• Scatter plot: Alcohol vs. Volatile Acidity

• Clusters loosely corresponded to:

  • Low quality
  • Medium quality
  • High quality

• Significant overlap observed

Hierarchical Clustering

• Average linkage method
• Dendrogram showed two main branches
• Primary split based on alcohol content

Unsupervised results reinforced findings from supervised learning.

---

⚠️ Obstacles & Observations

Obstacles

1. Edit Domain confusion:

   • Without converting the target to categorical:

     • Regression metrics (MSE/RMSE) appeared
     • Naive Bayes error: “Categorical class variable expected”

2. SVM training was slow on ~5,000 instances.

3. Class imbalance made accuracy misleading.

---

Key Observations

• Alcohol content is consistently the strongest predictor.
• Orange’s visual workflow makes model comparison simple.
• CN2 Rule Induction performed surprisingly well.
• AUC is more reliable than accuracy for this dataset.

---

🚀 Suggestions for Improvement

1. Add a Normalize widget before SVM and kNN → Likely to significantly improve their performance.

2. Reduce class imbalance by binning quality:

   • Low: 3–4
   • Medium: 5–6
   • High: 7–9

3. Add Random Forest to the comparison (ensemble methods typically outperform single trees).

4. Use AUC or F1 score as the primary evaluation metric instead of accuracy.

---

📚 Tools Used

• Orange 3
• UCI White Wine Quality Dataset
• 5-Fold Stratified Cross-Validation