04Classifiers and decoders

Graph network toolkits (PyG, PyG Temporal, DGL)

The three libraries football graph models are built with. They supply message passing, batching and recurrent graph layers. The football has to come from you.

  • Tracking
  • Graph
  • Learned
  • Needs labels
Message passing: each player updates from his neighbours along the edges, then a readout answers.
Modality
A player graph per frame
Output
Node, edge or graph predictions, after training
Evidence
Standard research frameworks; CIKM 2021 (PyG Temporal)
Code and weights
Frameworks only; no football weights
Licence
MIT (PyG, PyG Temporal), Apache-2.0 (DGL)

A football frame is a graph before anyone draws it. Players are nodes, and what matters lies between them. Graph neural networks compute on that structure by message passing: each node gathers information from its neighbours along edges, updates itself and repeats. A readout turns node states into a prediction about a player, a pair or the whole frame.

Library What it adds Licence
PyTorch Geometric The common layers (GCN, GraphSAGE, GAT, GINE) and batching of graphs of different sizes MIT
PyG Temporal Recurrent graph layers for sequences of graphs MIT
DGL The same ground with strong support for heterogeneous graphs Apache-2.0

PyTorch Geometric is the default. PyG Temporal matters because a match is a sequence of graphs. DGL earns its place when player, ball and goal are different node types.

Where it wins

  • Relations are explicit. Distance, closing speed and angle live on edges instead of being flattened into a vector.
  • Permutation invariance and variable player counts come with the formalism.
  • Edge-level outputs match football questions: who passes to whom, who presses whom.
  • Mature ecosystems. unravelsports builds football graphs for them directly.

What it costs

  • These are frameworks. Topology, features, architecture and labels are all design decisions, and each changes the result.
  • Graph models are hungry for data. Published football GNNs train on many matches of labelled tracking.
  • Temporal graph layers assume a node is the same entity across frames. Team-only tracking breaks that assumption.
  • They are GPU-oriented, with dependency stacks that complicate reproducible CPU builds.

How it plugs into possession

Graph networks are where supervised possession modelling is heading. PathCRF is one, and so are the phase models. The sound pipeline is a graph encoder for each frame, a sequence model across frames and a structured decoder at the end.

Mind the ball. The obvious graph has a ball node connected to every player. When the ball is untracked that node vanishes and the model must still answer, so the architecture needs a ball-free path from the start.

Two facts kept a graph network out of our Initial Challenge build: no labels on the match, and players that carry a team and no identity. Both calls are explained in Architecture, under Evidence, then a decoder and Players without names.

Our read

The right formalism for football, waiting on labels. Choose the graph before you choose the library.

Sources

  1. PyTorch Geometric repositorygithub.com
  2. PyTorch Geometric Temporal: spatiotemporal signal processing with neural machine learning models (CIKM 2021)github.com
  3. Deep Graph Library (DGL) repositorygithub.com