Graph representations for the analysis of multi-agent spatiotemporal sports data

Abstract

Analyzing tactical patterns in invasion games using multi-agent spatiotemporal data is a challenging task at the intersection of computer and sports science. A fundamental yet understudied problem in this area is finding an optimal data representation for processing athlete trajectories using machine learning algorithms. In the present work, we address this gap by discussing common representations in use and propose Tactical Graphs, an alternative graph-based format capable of producing integrative, contextualized models for machine learning applications. We provide an in-depth, domain-specific motivation of the proposed data representation scheme and show how this approach exploits inherent data traits. We propose Tactical Graph Networks (TGNets), a light-weight, hybrid machine learning architecture sensitive to player interactions. Our method is evaluated with an extensive ablation study and the first comprehensive state of the art comparison between standard feature, state vector, and image-based methods on the same dataset. Experiments were conducted using real-world football data containing short sequences of defensive play labelled according to the outcome of ball winning attempts. The results indicate that TGNets are on par with state-of-the-art deep learning models while exhibiting only a fraction of their complexity. We further demonstrate that selecting the right data representation is crucial as it has a significant influence on model performance. The theoretical findings and the proposed method provide insights and a strong methodological alternative for all classification, prediction or pattern recognition applications in the areas of collective movement analysis, automated match analysis, and performance analysis.