Wine Quality Prediction

Drew Colbert 2023-08-14

Project Summary

This project set out to determine if we could predict the ratings of a wine based on the chemical make up of the wine and certain intrinsic properties of the wine. To do this we found a data set containing ~6500 ratings of both red and white wine and 11 intrinsic properties of the wine. Each wine received a rating of 1-10, 10 being excellent and 1 being bad. In order to try and predict the rating, a randomForest model was trained and tested on the data with the goal of obtaining the highest test accuracy possible. Different transformations and inputs were tried with our randomForest model to try and boost the accuracy. I found that I was able to predict with 83% accuracy the rating of a wine. The biggest factor that determined whether or not a wine was rated highly was the level of alcohol in the wine. As the alcohol amount went up, so did the rating on average.

This finding was tested to see if it was just by chance or not. An independent, two-tailed t-test was run on the alcohol levels of wines rating ‘High’ (7-10) and ‘Low’ (1-6). I found that the average alcohol level in the ‘High’ rated wines was significantly higher than that of the ‘Low’ rated wines. With a difference in the average alcohol level being between 0.114 and 0.110 mL which was found to be significant with a p-value of 2.2 x 10^-16, suggesting that alcohol does play a big influence in the rating that a wine receives. A one way ANOVA was run to find if the mean alcohol level differed between individual rankings. It was found that ratings 9 and 8, 9 and 7, 4 and 3, 5 and 3, 6 and 3 were the only ones that did not show a significant difference in alcohol level. This coincides with our t-test as these ratings fall within the same ‘High’ or ‘Low’ rating and would not necessarily be expected to be very different.

Project Details

Data Preview

## 'data.frame':    6497 obs. of  13 variables:
##  $ fixed.acidity       : num  7.4 7.8 7.8 11.2 7.4 7.4 7.9 7.3 7.8 7.5 ...
##  $ volatile.acidity    : num  0.7 0.88 0.76 0.28 0.7 0.66 0.6 0.65 0.58 0.5 ...
##  $ citric.acid         : num  0 0 0.04 0.56 0 0 0.06 0 0.02 0.36 ...
##  $ residual.sugar      : num  1.9 2.6 2.3 1.9 1.9 1.8 1.6 1.2 2 6.1 ...
##  $ chlorides           : num  0.076 0.098 0.092 0.075 0.076 0.075 0.069 0.065 0.073 0.071 ...
##  $ free.sulfur.dioxide : num  11 25 15 17 11 13 15 15 9 17 ...
##  $ total.sulfur.dioxide: num  34 67 54 60 34 40 59 21 18 102 ...
##  $ density             : num  0.998 0.997 0.997 0.998 0.998 ...
##  $ pH                  : num  3.51 3.2 3.26 3.16 3.51 3.51 3.3 3.39 3.36 3.35 ...
##  $ sulphates           : num  0.56 0.68 0.65 0.58 0.56 0.56 0.46 0.47 0.57 0.8 ...
##  $ alcohol             : num  9.4 9.8 9.8 9.8 9.4 9.4 9.4 10 9.5 10.5 ...
##  $ quality             : Factor w/ 7 levels "3","4","5","6",..: 3 3 3 4 3 3 3 5 5 3 ...
##  $ wine_type           : Factor w/ 2 levels "red","white": 1 1 1 1 1 1 1 1 1 1 ...
## NULL

Distributions

Most of the values here are left skewed with very long tails to the right. This tells me that there could be some outliers that may influence the ratings. Transformations can be made on the data that can reduce the amount of left skew that is seen in this data.

Correlations

All boxes highlighted in grey are relationships with a correlation value of 0.40 or greater, indicating the correlation is noteworthy. Alcohol and density are highly correlated with a value of -0.69. However, the most important one that relates to our goal is alchol having a correlation of 0.44 to the quality. This is the only factor that is this significantly correlated to the quality, followed by the volatile acidity at -0.27. This will be something to keep an eye on when we are making our predictions.

Predictions

Our predictor variable is quite imbalanced. Most of the ratings are 5 or 6, this will be something that will need to be addressed for predictions. Even the groups are highly imbalanced.

From our correlation matrix, we saw that alcohol was the most correlated to the quality. Here we see a very interesting pattern. It is clear that the the average amount of alcohol does go up as the quality goes up. This is good, this means alcohol will probably be the main factor in determining the quality of a wine.

Modeling

I will be creating 10 different randomForest models, with different transformations and input variables and comparing the accuracy of each. The following models will be made:

Default: No transformations or changes to the input variables
Log Transformed: Our inputs were skewed left, by taking the log of each input, we make them more normally distributed
Normalized: Z-score normalization is performed on the inputs before training
More Trees: Increased the number of trees in the model to 1500 instead of 500
Simple: Use less input variables, only use the top 5 inputs that effect the outcome the most
Balanced: Use over and under sampling techniques to balance out the number of observations for each group to 500
Combination: The data will be normalized and only the top 5 input variables will be used
XGBoost: Multi softmax XGBoost will be used to predict the quality level
Grouped: Instead of predicting the exact rating, predict the group (High, Medium, Low)
Balanced Grouped: Use over and under sampling on the groups to make 1000 observations in each group

Results

The Balanced, Balanced Grouped, and XGBoost all performed the best with test accuracy levels in the mid to low 80s. The grouped randomForest would not be the best choice to use because we lose the precision on our predicitons and we only gain about 4% of accuracy, which to me is not worth it. So the XGBoost model or the balanced model would be the best and would yield the best results. The balanced model runs quicker than the XGBoost does, so that would be the best one to use due to the better performance.

The goal of the company is to make wines that are rated at least a 7 or above. But these ratings can be so subjective and it could be that these wines were rated 7 or above by chance, and the amount of alcohol had nothing to do with it. To test this we are going to split the data into a ‘High’ rated group (7-10) and a ‘Low’ rated group (1-6). A t-test will be used to compare the means of these groups. Here we can see the distibutions for each group, right away we can see there could be a difference, but lets look further to confirm.

The results of our t-test show that we can reject the null hypothesis, and we can say that the difference in the average amount of alcohol between the ‘High’ rated group and the ‘Low’ rated group is not 0. There is a significant difference in the means of each group, and running this test 100 times guarantees the actual difference in the mean falls within 0.1008 and 0.1141 95 times.

To further explore and become more precise, an ANOVA was run on each individual rating to see if differences were found within and between each rating. The ANOVA resulted in an F-statistic of 320.4 and a coinciding p-value of 2 * 10^-16. Since the ANOVA was significant, a TukeyHSD test was implemented and plotted below to find out exactly which groups have a significant difference and which ones do not. The bars colored red means that the null hypothesis was not rejected and there is not a difference in the mean alcohol levels.

What we see is that only those wins that are closely rated (9-8, 9-7, 4-3) are the ones are not significant.

← Previous Post