Deep Learning with Spatiotemporal Data for Team Sports Analysis

Team sports analysis is crucial for enhancing performance in the global sports market. A
recent game changer in this respect has been the widespread acquisition of data, such as
spatiotemporal data from player tracking. However, extracting meaningful information on
collective movements from this data is a difficult task. Previous studies have thus begun to
utilize deep learning algorithms due to their high predictive power. Although these studies
propose isolated applications, a systematic investigation on the usage of deep learning for
analyzing spatiotemporal data in team sports is lacking.
The present thesis addresses this gap through a series of peer-reviewed publications. First,
theoretical considerations on spatiotemporal data in team sports analysis are discussed,
leading to the proposal and evaluation of a neural network architecture based on graph
representations of spatiotemporal sports data. Numerical results reveal that graph representations
and corresponding deep learning models can achieve state-of-the-art performance
in sports-related prediction tasks by leveraging the characteristics of spatiotemporal
data, while being computationally more efficient than comparable architectures.
Subsequently, several team sports analyses are conducted using the proposed network.
These analyses focus on the relationship between machine and human performance in
label prediction as well as on task and model designs that promote robust applications and
analyses previously deemed infeasible. Results show that deep learning models are well
suited for automation tasks such as supervised label generation. Yet, as the ambiguity
in ground-truth labels increases, sound operationalization and careful interpretations of
findings are required. If these conditions are met, deep learning is capable of producing
results that open new frontiers for team sports analyses.
The empirical work of this thesis is complemented by contributions addressing technological
challenges identified along the way. For example, profound programming experience
is often required to conduct advanced team sports analysis. To meet these challenges,
extensive code bases and sample datasets are released under open source licensing to significantly
simplify such analyses.
It can be concluded that deep learning is a powerful tool for team sports analysis with
spatiotemporal data which outperforms previous approaches in the investigated cases. The
presented findings offer valuable insights on the effective use of deep learning with respect
to task and model design as well as result interpretation. Furthermore, the released code
bases facilitate fast and reproducible implementations in the area of team sports analysis
as demonstrated.