This is a Machine Learning project developed as a part of the course in Inteligent Systems at the Faculty of Organizational Sciences, University of Belgrade, Serbia.
There are several ways to name player positions in basketball, but which one is the best? In the domain of basketball, usually one of three position classification schemes are used:
- Three Positions Classification - The traditional way of describing positions is:
- Guard (1)
- Forward (2)
- Center (3)
- Five Positions Classification - Today the most common approach is to use five positions:
- Point Guard (1)
- Shooting Guard (2)
- Small Forward (3)
- Power Forward (4)
- Center (5)
- Nine Positions Classification - There is also an alternative approach with players who can play at different two positions:
- Point Guard (1)
- Combo Guard (12)
- Shooting Guard (2)
- Guard Forward (23)
- Small Forward (3)
- Point Forward (34)
- Power Forward (4)
- Forward Center(45)
- Center (5)
Following is a visual representation of the position naming methods:
In this project, a machine learning approach is used to predict player's position based on his individual statistics, such as: three points made, blocks, games played, height etc. And a goal of the project is to define the best position classification scheme for basketball players.
In order to accomplish this, results of three classification algorithms are compared:
- Naive Bayes
- Suport Vectore Machine (SVM) - used with a linear kernel
- Logistic Regresssion
#Data Collection Since there is no available and free API for collecting the necessary data for the analysis, a website RealGM Basketball was sraped. The website contains a rich archive with statistics for thousands of professional basketball players.
#Data Exploration The collected data is made out of 26 different features. Following is the DataFrame representation:
Feature names:
| Short | Long | Short | Long |
|---|---|---|---|
| GP | Games Played | GS | Games Started |
| MIN | Minutes Played | FGM | Field Goals Made |
| FGA | Field Goals Atempted | FG% | Field Goal Percent |
| 3PM | 3 Points Made | 3PA | 3 Points Atempted |
| 3P% | 3 Points Percent | FTM | Free Throws Made |
| FTA | Free Throws Atempted | FT% | Free Throws Percent |
| OFF | Offensive Rebounds | DEF | Defensive Rebounds |
| TRB | Total Rebounds | AST | Assists |
| STL | Steals | BLK | Blocks |
| PF | Personal Fouls | TOV | Turnovers |
| PTS | Points | YR | Years playing |
| POS | Positions | W | Weight |
| H | Height | NAME | Name |
Features refference http://basketball.realgm.com/info/glossary
The data is split into three datasets:
- Dataset 1 - For 3 positions classification (1269 instances)
- Dataset 2 - For 5 positions classification (712 instances)
- Dataset 3 - For 9 positions classification (1429 instances)
#Feature Selection
For each dataset, feature "NAME" was removed because it is not relevant for this classification problem. Four different feature selection methods were used on each dataset, so actually, each out of three datasets have four variations.
Feature selection methods:
- None - No feature selection was done (all features were used).
- Domain knowledge - Features are removed using experience, common sense and knowledge about the game.
- Univariate Feature Selection - Select Percentile - This helps determine features with the highest score/relevance for predicting the specified label.
- Variance Threshold - Removes all features that don't meet a specified threshold.
This project was implemented using Python 2.7 and the following third party libraries:
- BeautifulSoup is a Python library for pulling data out of HTML and XML files. It works with your multiple parsers to provide idiomatic ways of navigating, searching, and modifying the parse tree. In this case it was used for scraping player statistics.
- Sckit-Learn is an open source Python library for machine learning built on top of NumPy and Scipy featuring many classification, regression and clustering algorithms.
- Pandas is an open source, BSD-licensed library providing high-performance, easy-to-use data structures and data analysis tools for Python. DataFrames were the most useful datastructure for this particular project because they enabeled easy data manipulation and spreadsheet-like visualisation
- NumPy is an open source extension for Python, which adds support for large, multi-dimensional arrays and matrices, along with a large library of high-level mathematical functions to operate on these arrays. NumPy enables any algorithm that can be expressed primarily as operations on arrays and matrices to run almost as quickly as the equivalent C code.
The project was developed using iPtyhon Notebook, an interactive command shell which added flexibility to the workflow and a much better user experience for this purpose than a standard shell, IDE or editor.
Data scraping was performed by the nba_web_scraper.py script.
The code performing classification on each three datasets can respectively be found at:
Cross-validation was used in order to assess the accuracy of the models.
###Feature selection
-
None - all features were used
-
Domain knowledge - the following features were removed: GP, GS, MIN, FG%, 3P%, FT%, PTS, YR, POS, 3PM, FTM, FGM.
-
Univariate Feature Selection - following features were found to be the most important:
- OFF
- AST
- BLK
- W
- H
- Variance Threshold - 3 features were removed:
- FG%
- 3P%
- FT%
###Classification results (accuracy scores):
| Feature Selection | Naive Bayes | SVM | Logistic Regression |
|---|---|---|---|
| None | 0.81 (+/- 0.08) | 0.90 (+/- 0.04) | 0.89 (+/- 0.06) |
| Domain Knowledge | 0.87 (+/- 0.07) | 0.91 (+/- 0.05) | 0.90 (+/- 0.06) |
| Univariate | 0.89 (+/- 0.04) | 0.91 (+/- 0.04) | 0.89 (+/- 0.06) |
| Variance Threshold | 0.82 (+/- 0.07) | 0.90 (+/- 0.04) | 0.90 (+/- 0.05) |
###Feature selection
-
None - all features were used
-
Domain knowledge - the following features were removed: GP, GS, MIN, FG%, 3P%, FT%, PTS, YR, POS, 3PM, FTM, FGM.
-
Univariate Feature Selection - following features were found to be the most important:
- OFF
- AST
- BLK
- W
- H
- Variance Threshold - 3 features were removed:
- FG%
- 3P%
- FT%
###Classification results (accuracy scores):
| Feature Selection | Naive Bayes | SVM | Logistic Regression |
|---|---|---|---|
| None | 0.70 (+/- 0.15) | 0.82 (+/- 0.12) | 0.69 (+/- 0.15) |
| Domain Knowledge | 0.74 (+/- 0.09) | 0.82 (+/- 0.09) | 0.72 (+/- 0.11) |
| Univariate | 0.77 (+/- 0.11) | 0.81 (+/- 0.09) | 0.72 (+/- 0.08) |
| Variance Threshold | 0.71 (+/- 0.11) | 0.82 (+/- 0.10) | 0.70 (+/- 0.12) |
###Feature selection
-
None - all features were used
-
Domain knowledge - the following features were removed: GP, GS, MIN, FG%, 3P%, FT%, PTS, YR, POS, 3PM, FTM, FGM.
-
Univariate Feature Selection - following features were found to be the most important:
- 3PA
- 3P%
- AST
- W
- H
- Variance Threshold - 3 features were removed:
- FG%
- 3P%
- FT%
###Classification results (accuracy scores):
| Feature Selection | Naive Bayes | SVM | Logistic Regression |
|---|---|---|---|
| None | 0.50 (+/- 0.13) | 0.57 (+/- 0.10) | 0.53 (+/- 0.10) |
| Domain Knowledge | 0.54 (+/- 0.13) | 0.58 (+/- 0.10) | 0.54 (+/- 0.06) |
| Univariate | 0.54 (+/- 0.13) | 0.56 (+/- 0.10) | 0.53 (+/- 0.07) |
| Variance Threshold | 0.52 (+/- 0.12) | 0.58 (+/- 0.10) | 0.54 (+/- 0.09) |
#Conclusion
For each of the three datasets SVM performed better than Naive Bayes and Logistic Regression.
The most accurate prediction of player position based on individual statistics can be made when classification is performed in three positions (3 classes)
In the case feature selection using the Univariate and Domain Knowledge methods gave the best results for all three classifiers. SVM performed the best, Logistic Regression was slightly behind with domain knowledge and Variance Threshold as feature selection methods. The worst results for each classifier were made when no feature selection was utilized.
Classification results are decent, but not as good as when classifying in three positions.
In this case Naive Bayes performed the best when Univariate feature selection was utilized, on the other hand SVM performed the best when the other three methods were used. Logistic Regression gave the best results when Domain knowledge and Univariate feature selection were applied.
In this case when Domain knowledge was applied as a feature selection method the classifiers performed the best.
Algorithms performed the worst on this dataset. This can be explained by the fact that many players from the dataset who can play at two positions spend most of the time playing at one particular position. As a result, there is a lot of players who play at position 45 and were classified to position 5, or from position 34 classified to position 3 etc.
For instance, Pedja Stojakovic can be found in the dataset with a position 34 (scraped from the website) even though he plays position 3. Similarly, when using Naive Bayes, confusion matrix contains 21 players who play at position 12 (able to play both 1 and 2), but are classified to position 2.