TY - GEN
T1 - Transformer-Based Framework for Versatile Analysis of Events Data in Soccer
AU - Rovshitz, Aviv
AU - Puzis, Rami
N1 - Publisher Copyright: © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.
PY - 2025/1/1
Y1 - 2025/1/1
N2 - In sports analytics, machine learning models are used for players’ appointment analysis, strategic planning, etc. Multiple rating systems were developed to rank the player performance based on statistical measurements, machine learning, and deep learning analysis of actions and sequences. However, capturing the complex relationships between actions within open play sequences, affected by tactical game planning and team deployment, is a major challenge. Inspired by language models, we developed SoccerTransformer, a transformer-based deep learning framework tailored to soccer analytics to address this challenge. The proposed framework includes (1) self-supervised pre-training to capture generic in-game dynamics, (2) a downstream sequence classification task to predict the attack phase outcomes, and (3) a player rating system. Empirical results demonstrate that SoccerTransformer accurately captures player roles and predicts goals with F1 of 0.814 to 0.862 on previously unseen games. SoccerTranformer framework stands out by providing a greater correlation with market values than other state-of-the-art rating systems.
AB - In sports analytics, machine learning models are used for players’ appointment analysis, strategic planning, etc. Multiple rating systems were developed to rank the player performance based on statistical measurements, machine learning, and deep learning analysis of actions and sequences. However, capturing the complex relationships between actions within open play sequences, affected by tactical game planning and team deployment, is a major challenge. Inspired by language models, we developed SoccerTransformer, a transformer-based deep learning framework tailored to soccer analytics to address this challenge. The proposed framework includes (1) self-supervised pre-training to capture generic in-game dynamics, (2) a downstream sequence classification task to predict the attack phase outcomes, and (3) a player rating system. Empirical results demonstrate that SoccerTransformer accurately captures player roles and predicts goals with F1 of 0.814 to 0.862 on previously unseen games. SoccerTranformer framework stands out by providing a greater correlation with market values than other state-of-the-art rating systems.
KW - language model
KW - market value
KW - soccer
KW - sports analytics
KW - transformers
UR - http://www.scopus.com/inward/record.url?scp=105002033062&partnerID=8YFLogxK
U2 - 10.1007/978-3-031-86692-0_7
DO - 10.1007/978-3-031-86692-0_7
M3 - Conference contribution
SN - 9783031866913
T3 - Communications in Computer and Information Science
SP - 80
EP - 92
BT - Machine Learning and Data Mining for Sports Analytics - 11th International Workshop, MLSA 2024, Revised Selected Papers
A2 - Brefeld, Ulf
A2 - Davis, Jesse
A2 - Van Haaren, Jan
A2 - Zimmermann, Albrecht
PB - Springer Science and Business Media Deutschland GmbH
T2 - 11th International Workshop on Machine Learning and Data Mining for Sports Analytics, MLSA 2024
Y2 - 9 September 2024 through 9 September 2024
ER -