Analysis of contextualized intensity in Men's elite handball using graph-based deep learning

Area of Science:

• Sports Science
• Machine Learning
• Data Analytics

Background:

• Manual data annotation in sports is time-consuming and limits analysis scope.
• Automating performance analysis can unlock deeper insights from large datasets.

Purpose of the Study:

• Introduce FAUPA-ML (Framework for Automatic Upscaled Performance Analysis with Machine Learning) to automate sports performance analysis.
• Leverage graph neural networks to scale expert knowledge to large datasets.

Main Methods:

• Trained graph neural networks on manually annotated handball match data.
• Applied the best network to analyze 539 elite handball matches (2019-2022).
• Calculated performance metrics (distance, speed, metabolic power/work) for attackers and defenders.

Main Results:

• Achieved 86% balanced accuracy in contextualizing match data.
• Counter attacks are shorter, less frequent, and more intense than position attacks.
• Attacking actions were found to be more intense than defending actions.

Conclusions:

• FAUPA-ML accurately replicates expert knowledge for performance analysis.
• Enables large-scale, contextualized analyses previously infeasible.
• Facilitates future research on factors influencing performance in team sports.