This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2020.2992025, IEEE Access Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000. Digital Object Identifier XX.XXXX/ACCESS.XXXX.DOI Machine learning enabled team performance analysis in the dynamical environment of soccer S. KUSMAKAR 1 , (Member, IEEE), S. SHELYAG 1 , Y. ZHU 1 , (Member, IEEE), D.B. DWYER 2 , P.B. GASTIN 3 , AND M. ANGELOVA 1 1 School of Information Technology Deakin University, Geelong, Australia, Vic 3125, (e-mail: s.kusmakar@deakin.edu.au) 2 School of Exercise & Nutrition Sciences, Deakin University, Geelong, Australia, Vic 3125, (e-mail: dan.dwyer@deakin.edu.au) 3 La Trobe Sport Exercise Medicine Research Centre, La Trobe University, Melbourne, Australia Vic 3086, (e-mail: p.gastin@latrobe.edu.au) Corresponding author: M. Angelova (e-mail: maia.a@deakin.edu.au). This work was supported by DSI collaborative research grant RM35517 “Intelligent sensor processing for enhancing defence decision support”. ABSTRACT Team sports can be viewed as dynamical systems unfolding in time and thus require tools and approaches congruent to the analysis of dynamical systems. The analysis of the pattern-forming dynamics of player interactions can uncover the clues to underlying tactical behaviour. This study aims to propose quantitative measures of a team’s performance derived only using player interactions. Concretely, we segment the data into events ending with a goal attempt, that is, “Shot ”. Using the acquired sequences of events, we develop a coarse-grain activity model representing a player-to-player interaction network. We derive measures based on information theory and total interaction activity, to demonstrate an association with an attempt to score. In addition, we developed a novel machine learning approach to predict the likelihood of a team making an attempt to score during a segment of the match. Our developed prediction models showed an overall accuracy of 75.2% in predicting the correct segmental outcome from 13 matches in our dataset. The overall predicted winner of a match correlated with the true match outcome in 66.6% of the matches that ended in a result. Furthermore, the algorithm was evaluated on the largest available open collection of soccer logs. The algorithm showed an accuracy of 0.84 in the classification of the 42, 860 segments from 1, 941 matches and correctly predicted the match outcome in 81.9% of matches that ended in a result. The proposed measures of performance offer an insight into the underlying performance characteristics. INDEX TERMS Dynamical systems, network science, distribution entropy, football, Kolmogorov com- plexity, machine learning, performance analysis, Shannon entropy, support vector machines, soccer. I. INTRODUCTION I MPROVING comprehension of strategic performance and success in team competition is an important goal in sports science [1]. Data-driven methods can effectively overcome the subjective limitations (manual analysis) of the match and offer better results for football clubs. Quantitative analysis can provide players and coaches with such insight, by allowing them to improve their match and assessment of the event beyond what personal observation can accomplish [2]. Traditionally, methods of performance analysis push the study of one-dimensional and discrete performance indica- tors towards probabilistic and correlational approaches [3]. However, this results in somewhat limited functional infor- mation as it lacks the understanding of the player-to-player interactions that support the actions of players and overall team behaviour. It is reasonable to expect an analysis of such one-versus- one dynamics in team sports to be insufficient as multiplayer interactions are important in determining success and fail- ure [4]. Therefore, in order to quantify and explain per- formance, it has been advocated that performance analysis in team sports must also focus on the interactions between players that sustain the overall team behaviour [5], [6]. From the dynamical systems view, the understanding of how the co-ordination emerges from the interaction among the system components, that is, the player-to-player interaction, is the VOLUME 4, 2016 1