Aline Paes

Predicting soccer match outcomes is valuable for coaches and players engaged in strategic planning, for enhancing fan engagement, and for the economic, media, and journalistic activities that surround the sport. The task requires analyzing large volumes of data, including team statistics, player performance metrics, and historical results. Although player and team interactions intuitively play a pivotal role in determining outcomes, automated prediction methods based on statistical models and machine learning often overlook this dimension. This paper proposes modeling such interactions using graph representations and investigates heterogeneous graph neural networks for extracting latent interaction features. We further propose integrating these graph-based features with tabular features derived solely from past match outcomes to assess the contribution of interaction information to predictive performance. Specifically, we employ state-of-the-art heterogeneous graph neural network models to extract latent interaction features and fuse them with the observed tabular features to train an XGBoost classifier. We evaluate the approach on eight datasets drawn from the top leagues of England, Spain, France, Brazil, Germany, Italy, the Netherlands, and Argentina. Across the analyzed seasons, the classifier using heterogeneous graph-generated features outperformed the tabular-only classifier in a single year, while the classifier combining both feature sets achieved competitive performance throughout. Notably, however, the combined model did not consistently improve upon the tabular-only baseline in terms of Ranked Probability Score (RPS): in most seasons it matched rather than exceeded the baseline. These findings indicate that graph-based interaction features carry a complementary signal that can be incorporated without degrading predictive performance, while motivating further investigation into how this signal can be more effectively exploited.