04Physics and features
unravelsports
Turns tracking into graphs, pressing intensity and formations. The shortest path from a kloppy dataset to a graph neural network.
- Tracking
- Graph
- Library
- MPL-2.0
- Modality
- Tracking loaded through kloppy
- Output
- Graph datasets; pressing and formation features
- Evidence
- Documented open library with worked examples
- Code and weights
- Python package; no possession model included
- Licence
- MPL-2.0
unravelsports does the unglamorous middle step of football graph learning. It takes tracking loaded through kloppy and builds, frame by frame, the graphs a network trains on: players as nodes with position and velocity features, edges with distances and angles, a label attached to each graph.
Around that converter it ships tactical models that need no training: a pressing-intensity measure and formation detection.
Where it wins
- It removes the data wrangling between tracking files and a graph framework.
- Pressing intensity is a physically meaningful pressure signal between defenders and attackers.
- It is built on kloppy, so every provider kloppy reads is supported.
- MPL-2.0: usable inside permissive projects, with changes to its own files staying open.
What it costs
- It prepares graphs and trains nothing for you. A possession model still needs labels and a network.
- Pressing and formation describe behaviour around the ball. They do not say who has it.
- Graph features assume one node per player per frame. Missing players need a policy.
How it plugs into possession
In two places. As preprocessing for a supervised graph model of possession, it is the bridge between readers and the graph frameworks.
As a feature source, pressing intensity is a decent soft cue: the team applying pressure is rarely the team in possession. It is a correlate, and should be weighted as one.
Our read
The practical on-ramp to football GNNs, with a pressing measure worth borrowing on its own.