Clustering algorithm for formations in football games

In competitive team sports, players maintain a certain formation during a game to achieve effective attacks and defenses. For the quantitative game analysis and assessment of team styles, we need a general framework that can characterize such formation structures dynamically. This paper develops a clustering algorithm for formations of multiple football (soccer) games based on the Delaunay method, which defines the formation of a team as an adjacency matrix of Delaunay triangulation. We first show that heat maps of entire football games can be clustered into several average formations: “442”, “4141”, “433”, “541”, and “343”. Then, using hierarchical clustering, each average formation is further divided into more specific patterns (clusters) in which the configurations of players are different. Our method enables the visualization, quantitative comparison, and time-series analysis for formations in different time scales by focusing on transitions between clusters at each hierarchy. In particular, we can extract team styles from multiple games regarding the positional exchange of players within the formations. Applying our algorithm to the datasets comprising football games, we extract typical transition patterns of the formation for a particular team.