Rating Player Actions in Soccer

Uwe Dick; Maryam Tavakol; Ulf Brefeld

doi:10.3389/fspor.2021.682986

Rating Player Actions in Soccer

Front Sports Act Living. 2021 Jul 15:3:682986. doi: 10.3389/fspor.2021.682986. eCollection 2021.

Authors

Uwe Dick¹, Maryam Tavakol², Ulf Brefeld¹

Affiliations

¹ Machine Learning Group, Leuphana University of Lüneburg, Lüneburg, Germany.
² UAI Group, Eindhoven University of Technology, Eindhoven, Netherlands.

Abstract

We present a data-driven model that rates actions of the player in soccer with respect to their contribution to ball possession phases. This study approach consists of two interconnected parts: (i) a trajectory prediction model that is learned from real tracking data and predicts movements of players and (ii) a prediction model for the outcome of a ball possession phase. Interactions between players and a ball are captured by a graph recurrent neural network (GRNN) and we show empirically that the network reliably predicts both, player trajectories as well as outcomes of ball possession phases. We derive a set of aggregated performance indicators to compare players with respect to. to their contribution to the success of their team.

Keywords: graph networks; soccer; sports analytics; trajectory data; trajectory prediction.