Sports-Analytics

Finding trends in sports data using different Machine Learning models.

This project gave me useful experience in Data Cleaning from large CSV files with NaNs, and testing multiple different models to see which fits best.

knn_position.py

• Attempts to model NBA positions by a players' height and weight using K Nearest Neighbors Classification.
• Features used = Height, Weight (2)
• Highest recorded test accuracy = 63.7% (K = 11)
• (K = 3) test accuracy = 53.3%
• In this learning problem, higher values of K will fit the model better to the test data, but the resulting graph does not appear grouped as Data_with_Lables.png does. This is because C-F and F-G positions fall in multiple classification groups (ex. C-F's can be amongst Centers or Forwards). Therefore, lower values of K represent these "intermediary" positions better since they are more dispersed, but overall there are more "1-position" players in our Dataset, so more C, F, and G are classified correctly (this can be seen in KNN Graphs).
• KNN graphs before and after classification are plotted in the KNN Graphs folder.

kmeans_position.py

• Attempts to model NBA positions by a players' height and weight using KMeans Clustering as Classification.
• Features used = Height, Weight (2)
• By using KMeans to classify our data, it only makes sense to use 5 and 3 clusters. 5 clusters attempts to use KMeans to assign each position in our data to a cluster, but the overlap between intermediary positions FG and CF and our inherently inseperable data does not allow clustering to sucessfully classify our data.
• Using 5 clusters, we reach test accuracy = 45%
• Using 3 clusters, we reach test accuracy = 52%
• 3 clusters generalizes our test data better for the same reason that KNN = 11 generalizes our test data better. Assuming there are only 3 categories of positions, the data becomes more seperable as the intermediary positions become the decision boundaries.
• In our dataset, clustering does not prove effective since our data is inseperable. But if we add another feature to our feature vector, we might be able to seperate our data. But in our dataset, we only have 2 quantatative features height and weight. Therefore, there is no way to further seperate our data.
• KMeans graphs before and after clustering are plotted in the Clustering Graphs folder.

nba_parse_csv.py

• Cleans large NBA CSV files into simple, usable data for knn_position.py and kmeans_position.py modeling.
• Drops NaNs, converts strings heights to integers, etc.
• Avoids parsing through CSV files every test run.