TY - GEN
T1 - Contextual Evaluation of Individual Contributions from Pressing Situations in Football
AU - Lee, Minho
AU - Jo, Geonhee
AU - Hong, Miru
AU - Bauer, Pascal
AU - Ko, Sang Ki
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.
PY - 2026
Y1 - 2026
N2 - Pressing is a key tactic in football for disrupting opponents’ strategies and creating advantageous opportunities. However, existing studies often rely on static, rule-based definitions, overlooking the dynamic evolution of pressing sequences and each player’s contribution. To address these limitations, this paper introduces exPressV2, a novel framework that uses a physics-informed ‘Pressing Intensity’ metric to dynamically identify high-pressure events from continuous tracking data. Our spatio-temporal architecture, combining a Gated Recurrent Unit (GRU) and a Graph Attention Network (GAT), then analyzes the sequence leading up to a press to predict the probability of regaining possession. The proposed framework was validated using 7,800 pressing situations from 36 K League 1 matches and demonstrated superior performance to baseline models with an ROC AUC of 0.731. This framework offers practical applications that translate the model’s predictive power into objective tools for both individual player evaluation and tactical simulation. The source code is publicly available at https://github.com/leemingo/express-v2.
AB - Pressing is a key tactic in football for disrupting opponents’ strategies and creating advantageous opportunities. However, existing studies often rely on static, rule-based definitions, overlooking the dynamic evolution of pressing sequences and each player’s contribution. To address these limitations, this paper introduces exPressV2, a novel framework that uses a physics-informed ‘Pressing Intensity’ metric to dynamically identify high-pressure events from continuous tracking data. Our spatio-temporal architecture, combining a Gated Recurrent Unit (GRU) and a Graph Attention Network (GAT), then analyzes the sequence leading up to a press to predict the probability of regaining possession. The proposed framework was validated using 7,800 pressing situations from 36 K League 1 matches and demonstrated superior performance to baseline models with an ROC AUC of 0.731. This framework offers practical applications that translate the model’s predictive power into objective tools for both individual player evaluation and tactical simulation. The source code is publicly available at https://github.com/leemingo/express-v2.
KW - Deep Learning
KW - Football Analytics
KW - Graph Neural Network
KW - Machine Learning
KW - Pressing Analysis
UR - https://www.scopus.com/pages/publications/105029083346
U2 - 10.1007/978-3-032-15165-0_3
DO - 10.1007/978-3-032-15165-0_3
M3 - Conference contribution
AN - SCOPUS:105029083346
SN - 9783032151643
T3 - Communications in Computer and Information Science
SP - 35
EP - 47
BT - Machine Learning and Data Mining for Sports Analytics - 12th International Workshop, MLSA 2025, Revised Selected Papers
A2 - Rios-Neto, Hugo
A2 - Robberechts, Pieter
A2 - Van Roy, Maaike
A2 - Zimmermann, Albrecht
PB - Springer Science and Business Media Deutschland GmbH
T2 - 12th International Workshop on Machine Learning and Data Mining for Sports Analytics, MLSA 2025
Y2 - 15 September 2025 through 15 September 2025
ER -