1

Wine Quality Dataset

Individual project – Beginner group

Yujin Kim (01815358)

Contents

1. Introduction to dataset
2. Data preprocessing with visualization
3. Modeling
4. Evaluation

2

1. Introduction to Wine Quality Dataset

• The dataset is related to red variants of the Portuguese “Vinho Verde” wine.
• The dataset describes the amount of various chemicals present in wine and their effect on its quality.

​

Challenges

• Imbalanced classes

​

Goal

- Creating a classification model to predict wine quality

3

1. Introduction to Dataset

• Load the data

​

​

​

​

​

​

​

​

​

​

4

1. Introduction to Dataset

• Columns (Wine characteristics)
• 11 Input variables (based on physicochemical tests)
• 1 Output variable (based on sensory data)

​

​

​

​

​

​

​

​

​

​

​

5

2. Data preprocessing

• Check the dataset and check for null and missing values

6

2. Data preprocessing

• Detect the outliers
• Use IQR method of identifying outliers to set up a “fence”.
• Any values fall outside of this fence are considered outliers
• Minimum & maximum fence

​

7

2. Data preprocessing

• Detect outliers to drop

​

​

• Drop the outliers

​

​

• Result

: 1599 -> 1562 rows 37 rows are dropped

​

8

2. Data preprocessing with visualization

• Correlation of each values

9

2. Data preprocessing with visualization

• Visualize the correlation matrix with heatmap

10

2. Data preprocessing with visualization

11

2. Data preprocessing with visualization

12

2. Data preprocessing with visualization

• Correlation Heatmap with values

13

2. Data preprocessing with visualization

• Find less correlated columns
• Less than 0.1

​

​

​

​

​

​

​

• 3 columns detected

-> Residual sugar, free sulfur dioxide, pH

14

2. Data preprocessing

• Drop less correlated columns

​

​

• After that, we have 9 columns (12 -> 9)

​

15

2. Data preprocessing

• Re-scaling

- First, check the skewness of each column

16

2. Data preprocessing

• Re-scaling
• For fixed acidity & total sulfur dioxide

-> log transformation

• For chlorides & sulphates

​

17

2. Data preprocessing

• Re-scaling
• Finally, check the skewness of each column to see the difference

18

3. Modeling

• Creating a dataset -> divide into input data(x) and output data(y)

​

​

​

​

​

​

​

​

​

​

• As you can see, the quality dataset have imbalanced classes

19

There are much more normal wines then excellent or poor wines

3. Modeling Artificial Neural Network(ANN)

• Seed fix

- Seed function

: used to save the state of a random function, so that it can generate same random numbers on multiple executions of the code on the same machine or on different machines

​

• Define a loss function & optimizer
• use Cross-Entropy loss
• Adam for gradient descent with learning rate 0.001

20

3. Modeling – ANN1 model

• Define an Artificial Neural Network:

- ANN1 model

- 7 hidden layers

21

3. Modeling

• Train the model
• Loop over the data iterator, and feed the inputs to the network and optimize
• the following code collects the loss and accuracy calculated while training the model
• made list of train accuracy and train loss to plot the ‘Train Accuracy vs Loss’ graph

22

3. Modeling

• Test the network
• Also, made list of test accuracy and test loss to plot the ‘Test Accuracy vs Loss’ graph

​

​

23

3. Modeling

• Train and validate the network on the training data.

- At first, set the epoch to 1000

24

3. Modeling

• Plotting the Accuracy graph
• With ‘train_accu’ list and ‘eval_accu’ list from the above train and validation
• Make a plot that shows train and valid accuracy

25

3. Modeling

• Plotting the Losses graph
• With ‘train_losses’ list and ‘eval_losses’ list from the above train and validation
• Make a plot that shows train and valid losses

26

3. Modeling

• In order to see them all together, plot the subplots
• As the epoch is too large, it’s hard to see the tendency right away

27

3. Modeling

• Decreased the number of epoch into 300 by splicing each lists, we can see the ‘Train vs Valid Loss & Accuracy’ in one sight.
• Looks similar with the graph below

28

3. Modeling

• Both of them show Overfitting

29

3. Modeling

• From the AI Winter School, we have learned how to prevent overfitting on Deep learning.
• First, Decrease the model complexity (hidden layers)

30

3. Modeling – ANN2 model

• Decrease the model complexity (hidden layers)

- Decreased the hidden layer from 7 layers to 4 layers

​

31

3. Modeling – ANN2 model

• Decreased epoch to 300
• But the accuracy is still very low

32

51 -> 55

3. Modeling – ANN3 model

• Decrease the model complexity (hidden layers)

- Decrease more hidden layer from 4 layers to 2 layers

​

33

3. Modeling – ANN3 model

• The accuracy is still low

34

55 -> 60

3. Modeling – ANN3

• Since we know that the data is imbalanced, we need to see the F1 score as well
• F1 score should be used when the dataset is balanced
• Also, it has ability to provide reliable results for a wide range of datasets whether imbalanced or not.
• Accuracy on the other hand struggles to perform well outside of well- balanced datasets

35

0.363

3. Modeling – ANN4 model

• Giving weights
• To each classes to see if it works to increase the F1 score

​

36

3. Modeling – ANN4 model

• Unfortunately, did not work

37

0.363

3. Modeling – ANN5 model

• Divided the weights to 1 to see if there’s change
• But even got worse

38

0.275

3. Modeling – ANN6 model

• SMOTE oversampling
• As changing the weight did not work well
• Tried to oversample the classes that have small sample

​

39

3. Modeling – ANN6 model

• Could observe the number of samples increased

- But F1 score still very bad

40

0.363

3. Modeling – ANN7 model

• Stratified sampling
• Slightly Better!

41

0.45

3. Modeling – RandomForest Classifier Model

• Use RandomForest Classifier
• Gives its classification report

42

0.58

3. Modeling - Gaussian Naïve Bayes Model

• Use Gaussian Naïve Bayes
• Gives its classification report

43

0.55

4. Evaluation

44