A Novel Clustering Framework to Identify Team Playing Styles Within Australian Football

Abstract A novel clustering framework is proposed to identify playing styles in invasion sports, particularly Australian football, based on match transactional data. These data are segmented at a granular level, representing an intrinsic partitioning of the entire match period into short-time interval event data. Clustering based on such granular time-event data provides opportunities to identify team playing styles at different periods within a match. Identifying effective team playing styles over short time intervals can be crucial for improving overall match strategies. Due to the complexity involved in forming interpretable clusters from vast amounts of data, this research has not been attempted before. As a key component of the clustering framework, this paper introduces a composite clustering assessment index, enabling the selection of the optimal clustering algorithm and the appropriate number of clusters. The composite index integrates subject matter expert knowledge, ensuring that the resulting clusters are not only data-driven but also aligned with the goals of the analysis. The framework is applied to identify offensive playing styles within Australian football by clustering possession chain data of offensive phases from 414 matches during the 2021 and 2022 Australian Football League (AFL) seasons. A total of 27 features from 112,530 offensive possession chains were clustered, identifying five distinct offensive team playing styles: (i) ‘Ineffective Ball Movers’, (ii) ‘Uncontested Ball Movers’, (iii) ‘Direct Ball Movers’, (iv) ‘Handball Ball Movers’, and (v) ‘Kick Ball Movers’.