Predicting Football Match Outcomes with Graph Neural Networks and Large Language Models: The “Smart Guess Football” Project

Amid the excitement of the 2024 European Championship, TuGraph programmers created an interactive demo named “Smart Guess Football” to explore the feasibility of applying cutting‑edge artificial‑intelligence techniques—graph computing, graph neural networks (GNN), transformers and large language models (LLM)—to sports event analysis.

Using publicly available data from European clubs and national teams, they constructed a player‑relationship graph covering roughly 30,000 athletes, encoding collaboration (passes) and confrontation (duels) as edges. This graph serves as the backbone for downstream modeling.

Feature engineering incorporated positional embeddings (both on‑field hotspots and tactical roles), basic personal attributes (age, strength, agility, dominant foot), and higher‑order traits generated by prompting a large language model with raw statistics to obtain richer descriptors. Team‑tenure information was also added to capture tactical cohesion.

The model pipeline feeds these node features into a GNN, followed by a Transformer‑style multi‑head attention module that computes buffs from teammates and debuffs from opponents. A special [CLS] token aggregates the match‑level representation, which is then passed through an MLP to produce logits for multi‑class win/draw/loss prediction.

Evaluated on 2023 matches as a training set and 2024 matches as a test set, the approach achieved 53% overall accuracy (including draws) and 71% accuracy when draws are excluded, based on seven test games. The authors note that data sparsity and imputed averages limit the current ceiling.

Beyond prediction, the system outputs attention weights that visualize player cooperation and opposition patterns, and the results are rendered in the TuGraph graph database for interactive exploration. The authors suggest further data enrichment and model tuning to improve performance.