Szmida, Patrik Peter; Toka, Laszlo
This paper proposes a framework for evaluating player actions in Counter-Strike 2 using Temporal Heterogeneous Graph Neural Networks. The framework transforms match replays into dynamic heterogeneous graphs, capturing both spatial and temporal patterns, then predicts win probabilities at a frame-by-frame level with over 76% accuracy. Using Shapley values, the model explains fluctuations in win probability by attributing them to specific player actions, such as eliminations or grenade usage. This enables professional teams to analyze the contextual impact of actions, improving strategic decision-making and performance evaluation.