Figure 2
The flowchart is arranged into six stages labeled from top to bottom as follows: “Raw sensor data”, “Spatial encoding”, “Adaptive topology”, “Multi-scale processing”, “Feature aggregation”, and “Output features”. The flowchart starts with a first text box in the “Raw sensor data” stage labeled “Input Layer (Player States)”. A downward arrow from the first text box leads to the second text box in the “Spatial encoding” stage labeled “Position Encoding (Fourier Features)”. A downward arrow from the second text box leads to a first rectangle in the “Adaptive topology” stage labeled “Dynamic Graph Construction” that consists of the third, fourth, and fifth text boxes labeled “Node Features: [position, velocity, stamina, role]”, “Edge Computation: Tactical Relevance Score”, and “Adaptive Threshold: tau (game underscore tempo, density)”. A downward arrow from the first rectangle leads to a second rectangle in the “Multi-scale processing” stage labeled “Multi-Head Graph Attention (6 layers)” that consists of the sixth, seventh, and eighth text boxes labeled “Layer 1 to 2: Local interactions (5 m)”, “Layer 3 to 4: Group dynamics (15 m)”, and “Layer 5 to 6: Team coordination (full)”. Each layer consists of a dark dot surrounded by a shaded region. From left to right, the shaded region becomes bigger indicating “Increasing receptive field”. A downward arrow from the second rectangle leads to a third rectangle in the “Feature aggregation” stage labeled “Hierarchical Pooling” that consists of the ninth, tenth, and eleventh text boxes labeled “Player-level features”, “Gorup-level features (attack or mid or defense)”, and “Team-level features”. A downward arrow from the third rectangle leads to a twelfth text box in the “Output features” stage labeled “Tactical Representation”. On the right, from top to bottom the six stages are labeled as follows: “O (n)”, “O (n minus d)”, “O (n-squared)”, “O (n-squared minus d)”, “O (n minus d)”, and “O (d)”, respectively.Dynamic graph neural network architecture