Predicting Wine Quality for California Wine Institute

Project goals

• To predict the quality of wine while incorporating unsupervised learning techniques.

Project description

• We've been tasked to find drivers of wine quality for the California Wine Institute.
• We've been given 2 datasets, one for red wine and one for white wine and are variants of the Portuguese "Vinho Verde" wine.

Initial hypotheses

• Higher amounts of Alcohol present determines the quality
• Volatile acidity content in the wine will affect the quality
• Chlorides content in the wine will affect the quality
• Density has an affect on quality

Acquire:

• Acquire the data from Data.World as CSVs
• You will need to use the winequality_red.csv and winequality_white.csv

Prepare

• Combined red (1599 Rows / 12 columns) and white (4898 Rows / 12 columns) csv's into one dataframe into a variable.
• Shape after Merge: 6497 Rows, 12 Columns
• Null Check: No Null Values
• Created is_red column for feature engineering in modeling
• We narrowed the dataframe to 6 Columns to work with
  • Quality, Chlorides, Volatile Acidity, Density, Alcohol, Is_red

Explore:

QUESTIONS OF THE DATA

• How does volatile_acidity effect quality?

• How does chlorides content effect quality?

• How does density effect quality?

• How does alcohol content effect quality?

Modeling:

• Use drivers in explore to build predictive models of different types
• Evaluate models on train and validate data
• Select the best model based on accuracy
• Evaluate the test data

Data dictionary:

Feature	Dtypes	Definition
Volatile Acidity	float	Acidity measured in grams per liter
Chlorides	float	Chlorides in mg per liter
Density	float	Measured in grams per liter
Alcohol	float	Alcohol by volume
Quality	int	Quality is a target variable in the dataset, indicating the overall quality or rating of the wine
Is red	int	1 = red / 0 = white binary

How to Reproduce

• Clone this repo (both the CSVs)
• Save both csv's into local folder
• Run notebook

Key findings

• RandomForest was the best at very surface level tuning
• It became overfit with more hyperparameters
• Even with only 5 features, it performed very well

Takeaways and Conclusions

• Ended up using volatile acidity, chlorides, density, alcohol, is_red to predict quality in wines.
• After running 594 Models with various iterations and scalers we determined that the best model was:
  • Total of 3 Clusters
  • Scaled with StandardScaler
  • Hyperparameters:
    • n_estimators = 300
    • max_depth = 6
    • min_samples_split = 10
    • min_samples_leaf = 1
• Random Forest Classifier output train accuracy of .601 and validate accuracy of .556
• Test Data Ran through the model returned a test accuracy of .562

Recommendations

• Recommend running the models to predict quality by keeping the color types of wine split.
  • This could enable for a more fine tuned model in determining quality.