bledem / wine_quality_prediction

Predict wine quality

We analyze Wine Quality Dataset [1]. The main challenges:

• multicollinearity of predictors,
• non obvious linearity between predictors and target,
• imbalanced multiclass.

Get ready

1. Create a new conda environement
2. Download dependencies:

pip install -r requirements.txt

3. Install nb_conda to use your conda env in jupyter notebook.

Usage

Use jupyter notebooks. We have used a separate python script for grid search for computation speed. We have prepared functions for ease of evaluation in the notebooks ( evaluate.py and functional.py).

Results

01 - multiclass

Predict wine quality with white wines as it is (from quality 3 to quality 9), as continous or categorical target. Best classifier:

• Random Forest
• XGBoost with all features.

ROC and PR curves for multiclass

02 - binary_07

Predict wine quality as binary target. We use the threshold t=7 to binarize the wine quality. Best classifier:

• Random forest
• XGBoost used with SMOTE resampling technique.

Macro f1 score result for different models and datasets

03 - binary_08

Predict wine quality as binary target. We use the threshold t=8 to binarize the wine quality. Best classifier:

• SVM
• MLP used with SMOTE resampling technique.

Macro f1 score result for different models and datasets

References

(P. Cortez, A. Cerdeira, F. Almeida, T. Matos and J. Reis. Modeling wine preferences by data mining from physicochemical properties. In Decision Support Systems, Elsevier, 47(4):547-553. ISSN: 0167-9236. (pdf)