Team sports, such as soccer and basketball, are mostly driven by the coordinated efforts of their players, leading to a complex system of interactions. Motivated by the availability of data related to these sports, recent methods were introduced to model the behavior of players using tracking data. While they produce convincing models of how players behave during matches, relations between them are left to, at best, be implicitly taken into account. For the last few years, neural networks have been central to state-of-the-art models on audio, image, and text data, which has prompted researchers to promote their use on other domains, such as graphs. In this research, we aim to use dynamic graphs as a representation for both soccer tracking data and interactions between players over time, designing new neural network architectures that can learn from this data. We aim to tackle two problems with this approach. The first is modeling of collective movement of players, with the intent of being able to realistically predict where players might be in the future. The second is the retrieval of match clips, that is, temporal ``slices'' of matches described by dynamic graphs. Our methodology employs a neural network capable of learning from dynamic graphs, under the Graph Network (GN) framework. This approach leverages the flexible formulation GNs provide, enabling us to explore different solutions within the same framework, while using the richer representation of a match with dynamic graphs.
News published in Agência FAPESP Newsletter about the scholarship: