Purnima Pathak, Anthony Pino, Emily Washburn, Sergio Roberto de Souza Junior
Tasty Wine manufacturing company wants to enhance their taste of wine and supply it to European countries to expand their business. They sought the help of Team-3 Agency to analyze the dataset of wine quality that has 6449 wine Ids. Based on the quality ratings in the dataset, Team-3 Agency’s task is to provide Tasty Wine with the hidden trends and patterns in what makes a good wine based on its ingredients so that they can manufacture quality wine for export. The Agency is using Unsupervised Machine Learning to find what ingredients affect the quality of wine.
The first thing we did was to clean the data. We used SQL Alchemy, created an engine that can connect to the database and queried the data and created a data frame. We dropped duplicate items and dropped unwanted columns. Then started the data exploration where we started the comparison between density and other wine ingredients like- fixed acidity, citric acid, chlorides, total sulfur dioxide, ph. value, alcohol content, volatile acidity, free sulfur dioxide, residual sugar, sulphates and quality.
Through our data exploration we found that 4 areas are seemingly affecting quality-
The lower the volatile acid the higher the wine quality. Volatile Acidity is associated with the gaseous acids in the wine. It is the smell and taste of vinegar in the wine.
The lower the chlorides the higher the wine quality. Chloride contributes to the saltiness in the wine.
The higher the citric acid, the higher the wine quality. Citric acid increases the acidity in the wine and gives it a “fresh” flavor.
The higher the alcohol content, the higher the wine quality. The alcohol content in the wine decides whether the wine is strong or not.
We went for PCA Method to do the unsupervised machine learning with our dataset. We checked the variation of wine density which came out to be 0.05187. The variation in wine density is very small and most likely it does not have any effect on the quality of wine, so we dropped the density column. The variation in alcohol content came out to be 6.9, so we are keeping it. We used pd.get_dummies on types which were red wine and white wine to turn it from categorical values to numerical values and we appended it to the main data frame and then dropped the type column. Fit transform method was used through scikit-learn library before using the machine learning algorithms to reduce noise, remove redundant features so that computational efficiency can be improved. We used Kmeans clustering algorithm and got the elbow curve and predicted and plotted the model.
We did a comparative study on how the clusters would look like if we removed quality from the Kmeans. Below are the plots-
• Prior to completing any unsupervised learning, we found that 4 data points, volatile acidity, chlorides, citric acid and alcohol content were the best indicators for wine quality.
• Additional analysis showed that cluster 0 was the lowest quality and 1 was highest quality on average.
• When comparing the clusters against the initial bar graphs we found that the clusters matched, signifying a correlation between the 4 data points and wine quality.
• Our lowest average quality cluster, cluster 0, had more volatile acidity and chlorides, on average than cluster 1, our highest average quality cluster.
• Cluster 0 also had a lower average alcohol and citric acid content, which also matches up to our initial findings.
• The KMeans algorithm was able to cluster wines that match up to the original dataset without using the quality column.
• Cluster 2 ended up being in between the two other clusters.
One could, in theory, plug the stats of their wine in this model, see which cluster it falls into, and infer whether their wine is of higher quality or not.
The model performed better without using the quality scores as a factor. Further testing with different datasets could be done to see how the model performs, though from our findings, we think this could help someone make better wine choices.
We do not know what geographical locations were taken into count when this dataset was created and we do not know what age group rated the wine quality. The other Limitation could be that it’s one dataset and we didn’t test any other one. There also isn’t much metadata on what the numbers mean, meaning like if it’s alcohol percentage or not, for example.
https://www.kaggle.com/datasets/subhajournal/wine-quality-data-combined
https://seaborn.pydata.org/archive/0.11/generated/seaborn.distplot.html
https://faculty.washington.edu/yenchic/18W_425/Lec6_hist_KDE.pdf