By Mackenzie Mitchell
The purpose of this project was to classify strains of cannabis as indica, sativa, or hybrid using a classification model. The data was obtained from The Strain API and from web scraping WikiLeaf. I was able to classify cannabis strains based on their effects with accuracy of 68.6% using a Logistic Regression one-vs-all model.
Dummy Baseline Model
Classification models with all features
Classification models with hand selected features
Classification models with PCA prinicpal component variables
Grid Search
The goal is to predict all indica strains as either indica or hybrid, and to predict all sativa strains as either sativa or hybrid. This is because of the significant difference in effects and results from using these two different types of strains. While indica is better for night use and to relax, sativa is better for day use. Therefore, if someone wanted to consume sativa but instead consumed indica, they would be very upset as they would become tired and lazy during the time they may have wanted to be creative and uplifted. Additionally, of course, the overall goal is to predict as many of the observations as the correct strain as possible.
The goal of this project was intially to classifiy a cannabis strain as indica, sativa, or hybrid based on it's chemcial components. After webscraping, I was only able to get about 500 observations. This was insufficient for my objective. Changing gears, I found The Strain API (http://strains.evanbusse.com/). Using this tool, I was able to gather almost 2000 observations with 33 features scraped from this API.
The features included the various effects of the cannabis strain as well as the thc content. The thc content of each strain was obtained from webscraping https://www.wikileaf.com/strains. Once joining the dataframes, I was left with about 1300 observations and 37 features, as I also engineered features that represented scores of positive effects, negative effects, and medical effects for each cannabis variant observation.
There did not seem to be class imbalance at first as no one class contained half the data, however, if I were to go back I might SMOTE or use another method in order to treat the fact that there were 1004 observations of hybrid strains and only 966 observations split between indica and sativa strains within my dataset.
The THC content seemed to be pretty stable across all classes, with no strain type having significantly more or less THC than the others. There are some outliers in the indica class with higher levels of THC while the other two classes do not experience these outliers that have these excess levels of THC.
In looking at how features change across the different classes, I was able to hand select certain features that seemed to distinctly change across the three classes. The selected features are shown as follows:
After implementing PCA and running models, the results actually seemed to decrease in their performance. However, this was great experience for me to work with PCA as I was confused about this concept and how to implement it prior.
The most populated class is hybrid, so predict all strains as hybrid. Accuracy: 57.85% Average F-1 Score: 0.7
This model improved accuracy from the baseline, but was not the best model. Logistic Regression with PCA components reduced accuracy to 57.85% Accuracy: 68.6
This model improved accuracy from the baseline, but was not the best model. Accuracy: 66.86% Average F-1 Score: 0.6
Accuracy: 68.31% Average F-1 Score: 0.6
Accuracy: 69.48% Average F-1 Score: 0.63
