Wine Quality

Miniproject for NTU SC1015

Group 9, Lab Group SC2

Zhang Danxu • Lohia Vardhan • Sannabhadti Shikha Deepak

---

Overview 💻

This miniproject focused on UCI Machine Learning Repository Wine Quality Data Set, trying to use various properties of a wine to predict its quality.

Libraries 📚

• numpy
• pandas
• matplotlib
• seaborn
• sklearn
• tabulate
• torch

File Structure 📂

• data/: datasets

  • raw/winequality-red.csv: raw dataset downloaded from UCI Website
  • refined_wine.csv: dataset after cleaning
  • reclassified_wine.csv: dataset after reclassification
  • predictors.csv: cleaned and scaled predictor variables
  • response.csv: cleaned response variable
  • winequality.txt: description for datasets

• src/: jupyter notebooks and python scripts

  • data-clean-up.ipynb: notebook for data cleaning;
  • eda.ipynb: notebook for exploratory data analysis
  • ml.ipynb: notebook for machine learning;
  • SGD.py: script for Stochastic Gradient Descent;
  • loss_history.png: learning curve of SGD, number of iterations w.r.t. loss function

• .gitigore: files to be ignored by git

• LICENSE: lincensing information

• README.md: basic information about our project

Data Clean-Up 🧹

In data clean-up, we first dropped null values, duplicated rows, and outliers. Since our data was imbalanced, we reclassifed our response variable quality to make it more balanced. Then we did feature scaling to make our predictor variables of similar scales.

Exploratory Data Analysis 🔎

Exploratory data anaysis includes histogram with kde plots, boxplots for data before and after reclassificatin, finding corrleation between variables, and preforming Point Biserial Correlation.

Machine Learning 🤖

In machine learning, we first used decision trees, but since we had a relative large number of features, decision trees was a bit overfitting. The second model we built was stochastic gradient descent, with PyTorch's BCE loss funcition and SGD optimizer. SGD used a simple linear layer, but there could be some non-linear relationships. So, the last model we built is support vector machine with RBF kernel. The final training accuracy for SVM is 90%, and testing accuracy is 89%.

Contributors 👨‍💻

• Sannabhadti Shikha Deepak: Data Clean-up
• Lohia Vardhan: Exploratory Data Analysis
• Zhang Danxu: Machine Learning