Machine Learning and Data Mining for Sports Analytics

Ulf Brefeld, Jesse Davis, Jan van Haaren, Albrecht ZimmermannCommunications 2024 · DOI · OpenAlex

This document represents the proceedings of the 10th International Workshop on Machine Learning and Data Mining for Sports Analytics (MLSA 2023). Rather than a single experimental study, it serves as a curated collection of the latest research at the intersection of data science and professional sports.

Summary

The MLSA 2023 proceedings showcase a maturing field where machine learning is moving beyond simple descriptive statistics into complex decision support for strategy, player valuation, and injury prevention. While soccer remains a primary focus, the workshop highlights a significant expansion into diverse domains like SailGP, table tennis, and rugby.

The Challenge: Actionable Interpretation

Despite the explosion of data in professional sports—ranging from high-frequency player tracking to detailed event logs—venues for rigorous, peer-reviewed sports analytics research remain relatively rare. The MLSA workshop aims to bridge the gap between the "science" of data mining and the "art" of sports management.

The core challenge addressed by the contributors is actionable interpretation: how can we transform raw, noisy data into insights that help a coach decide on a match tactic, a scout value a player, or a medical team prevent an injury? The workshop specifically addresses the increasing complexity of sports data, such as the need to synchronize disparate event and tracking data streams and the demand for models that can handle the temporal dynamics of team play.

As a collection of 18 selected papers, the volume covers a wide methodological spectrum. Key technical approaches highlighted include:

  • Temporal Graph Networks (TGNs): Used for predicting pass receivers and outcomes in soccer by modeling the evolving relationships between players on the pitch.
  • Masked Autoencoders (MAE): Applied to event classification in elite soccer, leveraging self-supervised learning to understand game patterns.
  • Rule-Based Synchronization: Novel approaches to align "event data" (what happened) with "tracking data" (where everyone was), a notorious hurdle in sports analytics.
  • Expected Metric Modeling: Expanding the "Expected Goals" (xG) framework into new territories, such as "Expected Turnover Danger" (xCounter) and "Expected Goals" for handball.
  • Clustering and Sequence Analysis: Used to identify player roles in ice hockey and signature movement patterns in rugby.

Data Sources

The research presented leverages a mix of proprietary and public data, including industry leaders like Statsbomb, Stats Perform, and Wyscout. Mentions of the NFL Big Data Bowl datasets and data from elite European football leagues are prominent, alongside unique datasets for niche domains like SailGP (sailing), table tennis, and professional cycling.

Key Insights and Ideas

While soccer has historically dominated, this edition saw a record diversity of sports. Notably, the workshop featured its first-ever paper on sailing analysis (SailGP), signaling that high-performance analytics is permeating even the most specialized sports.

Researchers are also moving away from valuing actions in isolation. New models emphasize that the value of a play is deeply dependent on the spatial and temporal context of the game. Furthermore, there is a clear trend toward building integrated Decision Support Systems (DSS) designed specifically to assist human decision-makers in basketball and other team sports.

The MLSA 2023 workshop was led by a diverse group of editors from Germany, Belgium, and France.

Workshop Results

The workshop received 31 submissions, of which 18 were selected (a ~58% acceptance rate) after a rigorous single-blind review process. Papers demonstrated high accuracy in predicting boat speeds in SailGP and identifying player roles in ice hockey. The preface notes that the implementation of strict financial fair-play regulations in European football is actively driving the adoption of these analytics tools to optimize player valuation and team spending.

The program committee included experts from both academia and industry (e.g., Zelus Analytics, SciSports, StatsPerform), reflecting the applied nature of the field.

Limitations and Future Outlook

The editors acknowledge that while the field is growing, it faces several hurdles. Much of the most valuable data remains behind the paywalls of private companies, which can limit academic progress. Additionally, future work needs to better integrate psychological and "mental" player evaluation into the currently physics-and-geometry-heavy models. Finally, the next frontier is moving these complex ML models from post-match analysis into real-time, sideline decision-making tools.

Summary updated

BibTeX
@inproceedings{brefeld2024machine,
  author = {Ulf Brefeld and Jesse Davis and Jan van Haaren and Albrecht Zimmermann},
  title = {{Machine Learning and Data Mining for Sports Analytics}},
  booktitle = {Communications},
  year = {2024},
  doi = {10.1007/978-3-031-53833-9},
}
Figure