Predicting NBA Player Chemistry Using Graph Neural Networks

Hello everyone, my name is sea_turt1e.

This article will share the process and results of building a machine learning model to predict player chemistry in the National Basketball League (NBA), a sport I love very much.

Overview

• Predict player chemistry using graph neural networks (GNN).
• The area under the curve (AUC) is used as the evaluation metric.
• The AUC at convergence is approximately 0.73.
• The training data covers the 1996-97 season to the 2021-22 season, and the data from the 2022-23 season is used for testing.

Note: About NBA

For readers unfamiliar with the NBA, parts of this article may be difficult to understand. "Chemical reaction" can be understood from a more intuitive perspective. Additionally, while this article focuses on the NBA, the method could also be applied to other sports and even interpersonal chemistry prediction.

Chemical reaction prediction results

Let’s look at the prediction results first. I'll go into more detail about the dataset and technical details later.

Explanation of sides and fractions

In chemical reaction prediction, red edges indicate good chemical reactions, black edges indicate moderate chemical reactions, and blue edges indicate poor chemical reactions.

The fraction on the side represents the chemical reaction score, ranging from 0 to 1.

Chemistry predictions for star players

Here are the chemistry predictions for star players. The graph only contains pairs of players who never played for the same team.

Looking at the predictions of star players who have never played together, the results may not always be intuitive.

For example, LeBron James and Stephen Curry showed excellent coordination in the Olympics, indicating good chemistry. On the other hand, Nikola Jokic is surprisingly predicted to have poor chemistry with other players.

Chemistry predictions for major trades in 2022-23 season

To bring the predictions closer to reality, I tested the chemistry between players in actual trades for the 2022-23 season.

Since data from the 2022-23 season is not included in the training data, predictions that match realistic impressions can indicate the effectiveness of the model.

There are several important trades happening in the 2022-23 season.

Here are the predictions for key players including Kevin Durant, Kyrie Irving and Rui Hachimura.

The chemistry predictions for their new team are as follows:

• Lakers: Rui Hachimura – LeBron James (red edge: good chemistry)
• Suns: Kevin Durant – Chris Paul (Black Side: Medium chemistry)
• Mavericks: Kyrie Irving – Luka Doncic (blue side: poor chemistry)

These results appear to be pretty accurate considering the dynamics of the 2022-23 season. (Though things changed for the Suns and Mavericks the following season.)

Technical details

Next, I will explain the technical aspects, including the GNN framework and dataset preparation.

What is GNN?

GNN (Graph Neural Network) is a network designed to process graph-structured data.

In this model, "chemical reactions between players" are represented as graph edges, and the learning process is as follows:

• Direct side: The pair of players with higher number of assists.
• Negative Side: A pair of players with a lower number of assists.

For negative edges, the model gives priority to “teammates with low assists” and weakens the influence of “players from different teams”.

What is AUC?

AUC (area under the curve) refers to the area under the ROC curve and is used as a metric to evaluate model performance.

The closer the AUC is to 1, the higher the accuracy. In this study, the AUC of the model was approximately 0.73—a middling to above average result.

Learning Curve and AUC Progress

The following is the learning curve and AUC progress during the training process:

Dataset

The main innovation lies in the construction of the data set.

To quantify chemistry, I assume "high assists" means good chemistry. Based on this assumption, the data set is structured as follows:

• Positive side: Players with high number of assists.
• Negative side: Players with low assists.

Additionally, teammates with low assist counts are explicitly considered to have poor chemistry.

Code details

All code is available on GitHub.

Following the instructions in the README, you should be able to replicate the training process and plot the graphs described here.

https://www.php.cn/link/867079fcaff2dfddeb29ca1f27853ef7

Future Outlook

There is still room for improvement and I plan to achieve the following goals:

1. Expand the definition of chemical reaction
   • Incorporate factors beyond assists to more accurately capture player relationships.
2. Improve accuracy
   • Improving AUC through better training methods and expanded data sets.
3. Integrated natural language processing
   • Analyze player interviews and social media posts to add new perspectives.
4. Write an article in English
   • Publish content in English to reach a wider international audience.
5. Developing GUI for graph visualization
   • Create a web application that allows users to interactively explore player chemistry.

Conclusion

In this article, I describe my attempts to predict NBA player chemistry.

While the model is still under development, I hope to achieve more exciting results with further improvements.

Welcome to leave your thoughts and suggestions in the comment area!

---

If you need further improvements, please let me know!

The above is the detailed content of Predicting NBA Player Chemistry Using Graph Neural Networks. For more information, please follow other related articles on the PHP Chinese website!