Representing StatsBomb 360 freeze frames as graphs (nodes = players, edges = relationships) and processing them through graph neural networks (GNNs) to predict possession outcomes. Unlike tabular approaches (which require hand-crafted features like "distance to nearest defender" and break when player counts vary) or image approaches (which lose structural information), graphs naturally handle variable numbers of players and preserve relational information. The model predicts three possession outcomes: shot, short possession (<5 seconds), or long possession (>5 seconds), enabling analysis of both attacking threat and defensive structure.
Graph construction: For each 360 frame, create a graph where:
Key design decision: encode the field of view as a global feature — the model learns that seeing 15 players in a narrow view means something different from seeing 2 players in a wide view. The model infers off-screen context from what it can see.
Training: Classify each possession state into shot / short possession / long possession. The model learns that high short-possession probability means the ball carrier is under heavy pressure with blocked passing lanes; high shot-possession probability means there are open passing options into attacking territory.
The bottleneck in tactical phase classification isn't model complexity — it's labeled training data. With GCN embeddings from 360 data, only ~500 manually labeled actions (about 30 minutes of analyst time) are sufficient to train a simple classifier that accurately labels ALL remaining actions into phases of play. The embedding captures the spatial structure; the classifier just needs a few examples of each phase. This 500-label approach is 100x more efficient than traditional manual video coding.