Project Title: Red Wine Quality Prediction and Deployment

Data Collection

• Sourced dataset from Kaggle: Red Wine Quality Dataset.
• Dataset contains 1,599 rows, 11 input features (e.g., fixed acidity, pH, alcohol) and 1 target variable (quality).

Exploratory Data Analysis (EDA) and Data Preparation

1. Data Formatting: Verified column data types and statistical descriptions using .describe().
2. Missing Values: Found no missing data, requiring no imputation or removal.
3. Duplicates: Identified and removed 240 duplicate rows, retaining only the first occurrence.
4. Outlier Handling: Used Interquartile Range (IQR) to detect and remove outliers with custom functions.

Feature Engineering

1. Feature Selection:
   • Created a correlation matrix and heatmap to analyze relationships between features.
   • Dropped low-impact features ('pH', 'fixed acidity', 'citric acid', 'free sulfur dioxide') based on their correlation with the target variable.
2. Feature Scaling: Applied MinMaxScaler to standardize data, enabling effective performance for Euclidean distance-based models.
3. Target Encoding: Binarized wine quality scores (Good: >7, Bad: ≤7).

Data Balancing

• Addressed class imbalance (125 "Good" vs. 860 "Bad") using SMOTE (Synthetic Minority Oversampling Technique) to ensure balanced training data.

Model Development and Evaluation

1. Train-Test Split: Divided data into 80% training and 20% testing sets.

2. Model Selection: Compared five classifiers using accuracy scores:

   • Random Forest (93.8%)
   • XGBoost (94.4%)
   • K-Nearest Neighbors (94.7%)
   • Decision Tree (90.1%)
   • SGD Classifier (84.6%)
   • Selected XGBoost as the best performer through 10-fold cross-validation with a mean accuracy of 92.8%.

3. Hyperparameter Tuning: Optimized parameters (learning_rate, max_depth, min_child_weight, gamma) using RandomizedSearchCV and GridSearchCV for peak performance.

4. Performance Metrics:

   • Precision: 0.91
   • Recall: 0.95
   • F1 Score: 0.93
   • Test Accuracy: 94.5%
   • Confusion Matrix:

     [619  66]  
     [ 32 659]
     

Model Deployment

1. Model Pickling: Saved the trained model as a pickle file and verified functionality with dummy scaled inputs.

2. Flask Web Application:

   • Developed a user-friendly webpage with a form for inputting wine parameters.
   • Validated inputs using Python's try and except for error handling.
   • Scaled user inputs and fed them into the model to predict wine quality (Good/Bad).
   • Displayed prediction results with a help table for parameter guidance and contact information.

3. Deployment on Render:

   • Created requirements.txt and Procfile for dependency management and app configuration.
   • Uploaded codebase to GitHub and deployed the app on Render, enabling public accessibility.

Impact

This project automates wine quality prediction, providing accurate and efficient insights for quality assurance. It integrates machine learning with a user-friendly interface, showcasing end-to-end machine learning project deployment.