NBA Clustering based on Similar Skillset

Background

• NBA players were grouped together based on their stats provided from: Basketball Reference

Code and Resources Used

• Python Version 3.7
• Jupyter Notebook, Spyder
• Basketball Reference

Files

• NBA_cluster.ipynb - Jupyter notebook for clustering NBA players based on similar skillset
• HC_Cluster.ipynb - Jupyter notebook for Hierarchical Clustering of NBA players based on similar skillset
• data_scrape.py - Python script to scrape NBA players stats
• nba.csv - csv file for all NBA players stats

The Dataset

• Scraping the data from Basketball Reference and manipulating it, the dataset contains:
  • NAME: Player's Name
  • Pos: Player's position
  • MP: Minutes Played
  • PER: Player Efficiency Rating
  • TS%: True Shooting %
  • 3PAr: 3pt Attempt Rate
  • Ftr: Free throw Attempt Rate
  • ORB%: Offensive Rebound %
  • DRB%: Defensive Rebound %
  • TRB%: Total Rebound %
  • AST%: Assist %
  • STL%: Steal %
  • BLK%: Block %
  • TOV%: Turnover %
  • USG%: Usage %
  • OWS: Offensive Win Shares
  • DWS: Defensive Win Share
  • WS: Win Shares
  • WS/48: Win Shares per 48 games
  • OBPM: Offensive Box Plus/Minus
  • DBPM: Defensive Box Plus/Minus
  • BPM: Box Plus/Minus
  • VORP: Value over Replacement Player
  • 2P: 2-Point field goal per game
  • 2P%: 2-Point %
  • 2PA: 2-Point Attempts
  • 3P: 3-Point field goal per game
  • 3P%: 3-Point %
  • 3PA: 3-Point Attempts per game
  • AST: Assists Per Game
  • BLK: Blocks Per Game
  • DRB: Defense Rebound per game
  • FG: Field goals per game
  • FT: Free throws per game
  • FT%: Free throw %
  • FTA: Free throw attempts per game
  • ORB: Offensive Rebound per game
  • PF: Personal Fouls
  • PTS: Points per game
  • STL: Steals per game
  • TOV: Turnover per game
  • TRB: Total rebounds per game
  • PPG: Average Points per game
  • Salary: Player salary

Clusters

• Utilziing the elbow method and silhouette scores, NBA players were grouped into 3 distinct clusters.

• Cluster 0 represents players who are generally more 'bigger' in the league and can grab rebounds

• We can see here, that cluster 1 favours more of the Fowards positions (C and PF)

• While cluster 2 favours the Guards positions (PG and SG)

• In an attempt to cluster each player by their positions (PG, SG, SF, PF, C) I ran the algorithm again for 5 clusters.

• We can see using 5 clusters, we were able to cluster each player respective to their position

• Clusters 3 and 4 represent players who play the Forward position (PF and C)

• Clusters 0 and 1 represent players who play the Guard Position (PG and SG)

• While cluster 2 is an interesting cluster as it clusters players who tend to have a higher shooting rate.

  • It clusters PF's who have a higher shooting percentage, e.g. Lauri Markkanen and Marc Gasol.

Anomaly Detection

• Using a Multivariate Gaussian Distribution, I was able to identify outliers within the dataset.
• The picture shown below highlights players who are anomalous, and a deeper understanding on why they are anomalous can be found within the jupyter notebook.