1 of 18

Effects of Wine Components on Quality

Introduction Data Science

Group 15 - 110522126 林亭延 111526006 翁培馨 111522015 陳萬鈿

1

2 of 18

Outline

2

Introduction
Data preprocessing
Experiment
Result
Conclusion

3 of 18

Introduction

3

4 of 18

Motivation

For wine, the quality of it will greatly affect its price, taste and smell.

There is a huge disparity in wine prices today and we wanted to use this project as an opportunity to understand why.

5 of 18

Data Features

1 - fixed acidity

2 - volatile acidity

3 - citric acid

4 - residual sugar

5 - chlorides

6 - free sulfur dioxide

7 - total sulfur dioxide

8 - density

9 - pH

10 - sulphates

11 - alcohol

12 - quality

6 of 18

Data preprocessing

6

7 of 18

Correlation

8 of 18

Percentage of Quality

Quality Data Discretization

3 ~ 4 = 0

5 ~ 6 = 1

7 ~ 8 = 2

9 of 18

Experiment

9

10 of 18

Random Forest

In the part of predicting wine quality, we used three algorithms: random forest, GBDT, and KNN.

Random forest is a model composed of many different decision trees.

It is a type of ensemble learning method,

where a group of weak models are combined to create a strong model. In the context of predicting wine quality,.

The random forest algorithm is known for being able to handle a large number of features,

and for being resistant to overfitting

在預測葡萄酒品質的部分，我們分別使用了random forest、GBDT、KNN三個算法來做測試。

首先是random forest，這是一個由許多不同的決策樹所組成的模型。

這個算法是從訓練集中抽取出 n 筆資料，再從這 n 筆資料隨機挑選 k 個特徵做樣本。

重複 m 次，產生 m 棵決策樹。最後採用多個決策樹的投票機制來得到最後的結果。

它的優點是每棵樹會用到的訓練資料及特徵都是隨機決定的，

並且與決策樹相比，不容易過度擬合。

11 of 18

GBDT

Gradient Boosting Decision Tree algorithm .
Each time the model is built, the gradient descent direction of the loss function of the model is built before.
Use the loss function to evaluate the model performance, the smaller the loss function, the better the performance.

GBDT is an algorithm that uses the gradient descent method to optimize the loss function of the model before each iteration.

The loss function measures the performance of the model, and a smaller value indicates better performance

The result of GBDT is an additive model composed of multiple decision regression trees, so the final prediction result of the entire model is the sum of the prediction results of all regression trees.

The loss function is used to evaluate the performance of the model. The smaller the loss function, the better the performance.

接下來是GBDT，這個算法是在每一次建立的模型，是在先前建立模型損失函數的梯度下降方向。

使用損失函數來評價模型性能，損失函數越小，性能越好。

GBDT的結果是一個由多棵決策迴歸樹構成的加法模型，所以最後整個模型的預測結果是所有迴歸樹預測結果的和。

GBDT算法的優點在於可解釋性強，可以自動做特徵重要性排序以及防止過擬和。

12 of 18

KNN

K-Nearest Neighbors algorithm.
Result is determined by a "majority vote" of k neighbors.
Depends on the distribution of neighboring points.
The larger the K is, the more noise can be tolerated, while the smaller the K is, the more it is affected by the surrounding abnormal factors.

The last is the KNN algorithm. It mainly determines the value of k first,

and then calculates the distance between the current data and the adjacent data,

Finally, find out the k nearest neighbors ,

check which group has the largest number of neighbors, and join that group,

The larger K is the more noise can be tolerated, and the smaller K is the more affected by surrounding abnormal factors.

The advantage is that KNN is simple and well understood ,it can be used on non-linear dataset.

最後是KNN算法。它主要是先決定 k 的值，接著計算目前該筆資料與相鄰資料之間的距離，

最後找出與自己最近的 k 個鄰居，查看哪組鄰居的數量最多，就加入哪一組，

而K越大，越能容忍噪聲，K越小，越受周圍異常因素的影響。�

優點的部分是算法簡單且易理解且可用於非線性的資料集。

13 of 18

Result

13

14 of 18

	Avg Precision	Avg Recall	Accuracy	RMSE
Random Forest	0.63	0.47	0.82	0.42
GBDT	0.58	0.49	0.81	0.43
KNN	0.48	0.38	0.81	0.44

15 of 18

Importance Factor for Quality

16 of 18

Conclusion

16

17 of 18

By analyzing the data of red wine, we have found that the Random Forest performed better than others.�
Using Random Forest as our best model for prediction, we determined five of the features as the most influential : alcohol、sulphates、volatile acidity、total sulfur dioxide and density.

18 of 18

Thank you for listening

18