Basketball is one of the sports in which statistical analysis is most applied, thanks to the huge amount of data that can be derived from any single game and the fact that statistical analysis is becoming more and more important for many professional teams.

The same 5 positions are used to describe basketball players since Mr. James Naismith invented basketball in 1891: Point Guard (PG), Shooting Guard (SG), Small Forward (SF), Power Forward (PF), Center (C).

In the context of basketball analytics, this work aims at finding a new way to define positions of basketball players.

Three main technologies have been used in this work: Self-Organizing Maps (SOM), Multi-Dimensional Scaling (MDS), Principal Component Analysis (PCA). SOM was the main component used to obtain the definition of new positions, while MDS and PCA were used to obtain a proper visualization of results.

Training set characterization

The reference dataset is the database of statistics of NBA players, referring to the 2015-2016 regular season.

The following features have been considered for the classification: Rebounds (TRB), Blocks (BLK), Assists (AST), Steals (STL), Turnovers (TOV), Personal fouls (PF) and Points (PPG).

A criteria used to create a proper training set was to include the 8 players with the highest value of average minutes played per game from each of the 30 teams of the NBA.

Implementation Architecture

The organization of the pipeline for processing the data is pretty straightforward. The software application was written using the Python language. It uses some external libraries have been integrated to provide the implementation of the required statistical concepts, namely SOM, PCA and MDS.

In particular, the PyMVPA library was leveraged to use the SOM algorithm.

A main module, called BasketballSOM, contains the main function that performs the whole SOM algorithm.
External modules have been implemented for the creation of the training set and the input set, and also for the visualization of results.

Results about classic positions

In this analysis, the SOM was trained with five elements, one for each classic basketball position: PG, SG, SF, PF, C. The output layer of the SOM is divided in 5 clusters, one for each position. Some atypical players have been mapped on this output layer to show how classic positions are not able to describe their way of playing anymore. Mapped players are LeBron James (LBJ23), Kawhi Leonard (KL2), Draymond Green (DG23) and Russell Westbrook (RW0).

Results with the new positions

Five new positions have been defined: 1) All-Around All Star, 2) Scoring Backcourt, 3) Scoring Rebounder, 4) Paint Protector, 5) Role Player.

Conclusions

The defined positions are satisfying in terms of description of players game. A more detailed analysis can be performed, considering more sophisticated statistical categories to see if better positions definition can be performed. The analysis should be extended to different leagues around the world, considering also non-professional players.

Related information

This work led to the publication of the paper

Bianchi, Facchinetti, and Zuccolotto, “Role revolution: towards a new meaning of positions in basketball”, in Electronic Journal of Applied Statistical Analysis (Special Issue STATISTICS IN SPORTS), Università del Salento, Vol. 10, No. 3, pp. 712-734, 2017.

The paper is open access and available for download.