Goal:
To implement web scraping to extract data and then evaluate both supervised (Linear Regression, Logistic Regression) and unsupervised machine learning classifiers (KMeans Clustering) in their success in predicting NBA 2K video game scores.
You can find a Medium post I wrote summarizing the web scraping portion of this project at this link: https://towardsdatascience.com/web-scraping-nba-2k-data-d7fdd4c8898c.
I web scraped data from basketball-reference.com (www.basketball-reference.com/leagues/NBA_2014_totals.html) and hoopshype.com (https://hoopshype.com/2015/11/05/these-are-the-ratings-of-all-players-in-nba-2k16/) to extract real life NBA and NBA 2k video game data between 2014 and 2016.
After scraping the NBA 2k ratings to a JSON format, I joined the data with the metrics of the real life NBA players. By right joining to avoid missing values, I only kept players that were included in the 2K game.
Both linear and logistic regression models gave similar root mean square errors of 7.57 and 7.76 respectively. The linear regression model gave a model score of 0.90 while logistic regression gave a much lower model score. This indicates a linear problem set with high correlation between metrics and 2K rating.
A KMeans Clustering model was used to cluster positions and explore if players truly made plays according to their position or if being a generalist was common. The optimal number of clusters was determined to be either 4 or 5, which is similar to the 5 core positions the NBA employs.
The results of the clustering analysis revealed significant intermixing between clusters when it cames to steals and defensive maneuvers, but less mixing for 2 and 3 pointers made. This might indicate less rigid defensive responsibilities for players.