Agenda

• Introduction
• Variables
• Feature Engineering
• Practical Questions to Address
• K-Means Clustering Analysis
• Clustering
• PCA’s
• LOF
• LOF Cutoff
• Conclusions
• Limitations

Introduction

• Dataset comes from Kaggle
• 2023-2024 regular season player stats per game
• Team changes are represented by duplicate player names
• Dataset was downloaded to Github

Variables

Feature Engineering

• 0 missing values

• Team: If a player played for multiple teams in the same season, they 1 observations per team, plus an additional observation where the Team variable is TOT. The TOT observations were kept and the other observations for the player were dropped

• Some variables had to be renamed due to limitations of R,as variable names cannot have symbols or start with a number:

Practical Questions to Address

• Can the dataset be split into separate groups using a multitude of different player statistics?

• Can we determine any players who are considered outliers in comparison to their group?

K-Means Clustering Analysis

• First step is determining the number of clusters (“K”) needed

• Below plots show that the ideal number of clusters is 2

Clustering

• Now that we know the amount of clusters needed, below are the clusters shown graphically

• Points scored was chosen as it serves as a great benchmark of player “effectiveness”

PCA’s

• Principle Component Analysis (PCA) is a tool used to reduce the dimensionality of the dataset. Currently there is 27 variables in the dataset, but this can be reduced to a smaller amount of PCA’s

• The below plot shows how many PCA’s are needed

• For a 90% cutoff, the number of PCA’s kept should be 8

LOF

• Local Outlier Factor (LOF) is an outlier detection algorithm that is used to identify outliers within their respective clusters given specific criteria

• The outlier criteria that I have selected is any player who has a combines z-score for the following variables higher than 3:

LOF Cutoff

• LOF uses a set amount of neighboring observations to determine outliers as well, so different cutoff amounts will produce different results

• Below shows AUC curves for LOF detection. The higher the AUC the better performing the model:

• Best k is 400, but this is very close to the number of observations in the dataset, so a lower number should be chosen

Conclusions

• Can the dataset be split into separate groups using a multitude of different player statistics?

• Yes! After determining the ideal amount of clusters, the dataset was split into 2 separate clusters

• Can we determine any players who are considered outliers in comparison to their group?

• Yes! Outliers were determined with predetermined criteria and identified through LOF procedures

Limitations

• These conclusions should only be applicable to Modern NBA players. Older NBA players played a different type of game, resulting in a different skewness of certain statistics. If a noticeable shift in the NBA game should occur, these conclusions should be under scrutiny and retested.

• These conclusions should not be used for any other basketball league (i.e. WNBA, NCAA), as each league has a different style of play that may be rendered obsolete