A0384
Title: Clustering of soccer possessions using a mixture of absorbing Markov point processes
Authors: Koffi Amezouwui - ENSAI (Ecole Nationale de la statistique et de l analyse de l information) (France) [presenting]
Brigitte Gelein - ENSAI (France)
Matthieu Marbac - CREST - ENSAI (France)
Anthony Sorel - Rennes 2 University (France)
Abstract: Motivated by the analysis and clustering of soccer game situations, a specific finite mixture of absorbing Markov point processes models is introduced to cluster soccer ball possessions based on their temporal, spatial, and categorical characteristics. Each possession is modeled as a marked point process that ends with a specific event that causes this loss of ball possession, such as a foul or an interception by an opposing player. Event types are modeled using finite-state Markov chains, inter-event times are assumed to follow Gamma distributions, and spatial displacements are represented by time-scaled truncated Gaussian distributions. Model parameters are estimated using a generalized expectation-maximization (GEM) algorithm. The model is validated through numerical experiments and applied to event data from professional soccer matches. The results reveal interpretable clusters that reflect distinct tactical behaviors. Meaningful possession patterns are extracted, providing insights into game dynamics for performance analysis and training.
