Project Description

Team sports rely on highly tactical and complex strategies often suggested by coaching staffs. For instance, basketball is a 5 vs 5 players' game where it is crucial for the team in offense to adapt a strategy that maximizes open space for the shooter. The defending team, on the other hand, will make sure to prevent such activity as much as possible. The same reasoning could be applied in Football where the sequence of passes aims to create space for a direct shot on target or potentially a decisive pass. Predicting the defensive strategy of an opponent team in response to a given offense could serve as a useful tool in deciphering that team's general behavior given their historical data. Early mathematical models relied on classical statistical approaches to perform trajectory prediction of the players. With recent advances in computing and data availability, machine learning techniques are being adopted to decipher more complex tactical patterns from videos, images and numerical data. For instance, Convolutional Neural Networks (CNN) have been used on Basketball players' tracking data for trajectory prediction in [1]. Although successful, such approaches assume euclidean geometry in their optimization metrics and would thus fail to account for the inherent data interactions at the team and player-level. Geometric deep learning on non-euclidean data is a promising approach whose potential remains under-explored not only in trajectory forecasting but also in sports analytics.

Methodology

Graph Neural Networks (GNNs) have been successful in numerous applications such as molecular research and traffic prediction [2]. The network-like behavior of team sports makes them interesting candidates for GNN applications. Early studies on sports analytics made use of graph-theory to quantify a team's dynamics by modeling the players as interconnected nodes [3]. In [4], Clemente et. al discuss how graph theory and social networks analysis can be applied to team sports. They explore several network metrics to be used in sports analytics , most importantly: Centrality types (Power, PagRank, Laplacian), Clustering coefficient, Network Density and Transitivity. Such pre-Deep Learning metrics are crucial to visualize team geometries and find patterns within their positioning data by studying their graph-like interactions.

In this project, we formulate the task of predicting the defensive strategy of a sports team as a graph neural network problem, hence combining the traditional aspects of graph theory with modern neural network solutions. To proceed, we aim to build a graph generative model that simulates a defense scenario given a certain offense.

Important note

Upon the student's request and depending on data availability for different types of team sports, this project is flexible with regards to the sport under study (football, basketball, etc) and the potential applications and information to be explored through GNNs.

References

1] M. Harmon, P. Lucey, and D. Klabjan, “Predicting shot making inbasketball learnt from adversarial multiagent trajectories,”arXiv preprintarXiv:1609.04849, 2016.

[2] L. Zhao, Y. Song, C. Zhang, Y. Liu, P. Wang, T. Lin, M. Deng, and H. Li,“T-gcn: A temporal graph convolutional network for traffic prediction,” IEEE Transactions on Intelligent Transportation Systems, 2019.

[3] J. H. Fewell, D. Armbruster, J. Ingraham, A. Petersen, and J. S. Waters, “Basketball teams as strategic networks,”PloS one, vol. 7, no. 11, p. e47445, 2012.

[4] F. Clemente, F. Martins, and R. Mendes, Social Network Analysis Applied to Team Sports Analysis. 01 2016