Wine Quality Analysis Project

Overview

This project develops machine learning models for wine data analysis using two public datasets:

• Red wine samples
• White wine samples

The work covers data analysis, preprocessing, model selection with hyperparameter tuning, and model export.

Objectives

The code implements three tasks:

1. Task 1 (Regression)

• Predict the alcohol value from physicochemical features.
• Models: Linear Regression and XGBoost Regressor.
• Metric: RMSE.

2. Task 2 (Binary Classification)

• Classify wine type (red vs white).
• Models: SVC and Random Forest Classifier.
• Metric: Accuracy and confusion matrix.

3. Task 3 (Multiclass Classification)

• Predict wine quality score.
• Models: SVC, Random Forest, Decision Tree, and a PyTorch neural network.
• Metric: Accuracy, confusion matrix, and classification report.

Repository Structure

• code/source code files/source_code_l31537.py: Full pipeline with Tasks 1, 2, and 3.
• code/source code files/qualidade_vinhos.py: Task 3 focused workflow.
• data/red_wine.csv and data/white_wine.csv: input datasets.
• code/trained_model_files: folder intended for serialized models.
• requirements.txt: Python dependencies.

Preprocessing Pipeline

• Merge red and white datasets.
• Add wine type column (Vinho: red=1, white=0).
• Fill missing values using column means.
• Remove outliers with quantile-based bounds and IQR filtering.
• Standardize independent variables before training.

Task 3 Model Training and Confusion Matrices

For Task 3 (multiclass quality prediction), four classification models were trained and evaluated with confusion matrices.

Best result: task3_RedeNeuronal.

SVC

Random Forest

Decision Tree

Neural Network (Best)

Environment Setup

1. Create a Python environment.
2. Install dependencies:

pip install -r requirements.txt

Running the Code

From the repository root, run either script:

python "code/source code files/source_code_l31537.py"

Use the terminal menu shown by each script.

Notes

• Task 3 neural network training is computationally expensive due to K-Fold and high epoch count.
• Model export filenames are currently written in the working directory unless manually changed in code.
• Consider moving all model outputs to code/trained_model_files for consistency.