Pass receiver selection is a fundamental task in football analytics, aiming to predict the intended receiver given a specific game state. The data are provided as event‑centered freeze‑frame observations, offering only partial player visibility without complete trajectories or stable identities, which forces the model to reason over anonymous visible candidates, opponent pressure, and recent context. To tackle this, we propose the Hierarchical Possession‑Aware Graph Pointer Network (HPGPN), formulating the task as variable‑size candidate prediction over visible teammates. HPGPN jointly models current player interactions, local event context, and possession‑level temporal dynamics. Specifically, it (1) represents the current pass situation as a graph where nodes are visible players and edges capture spatial interactions; (2) embeds fixed event context (e.g., match phase, field location) into a global graph feature; (3) incorporates dynamic possession history via a dual‑branch design to capture the evolution of the attacking sequence. Candidate representations are refined hierarchically by integrating spatial, contextual, and historical evidence, and a glimpse pointer head scores the candidates to produce the final receiver prediction. Experiments on public football event and freeze‑frame datasets show that HPGPN significantly improves pass receiver selection performance. Ablation studies confirm the effectiveness of graph‑based interaction modeling, fixed event context, and the dual‑branch dynamic possession‑history modeling.
Review