05The engine

Evidence, then a decoder

Where should temporal consistency live: inside one model that reads many frames, or in a stage of its own behind the evidence? One learner, or two machines in a chain. The answer decides what can be tested, and what the design needs in labels.

  • One model or two
  • Decoder
  • Auditable
  • Fielded
The two-machine answer. Evidence flows into the decoder and also past it, into the audit, so every frame the decoder changed can be shown against the raw scores.
Option one
One multi-frame model: time lives in the weights
Option two
Evidence per frame, then a decoder over the half
What decides
The labels that exist, and what must be auditable
Contract
Three scores per frame between the stages

The design question

Should one model own both jobs, reading the evidence and keeping the sequence coherent, or should they be two machines? Why possession needs a sequence decision at all is set out in Why possession is a decoding problem. This study is about where that decision lives.

The two answers

  One model, many frames Evidence, then a decoder
Where time lives Inside the weights In a stage of its own
Needs Frame-level labels from other matches Scores per frame; labels optional
Guarantees Nothing about coherent spells The best path under stated costs
Inspect it by Its outputs only Each stage, and the difference between them
Breaks when Labels are scarce or from another provider A heavy decoder erases real short spells
The one-model answer. A window slides over stacked signals and one learner reads it. What it knows about time, it learned from labels, and nothing in it forbids a flicker.

If time lives inside the model

Input representation Learner What it captures
Engineered window features LightGBM, XGBoost Trends and time statistics, made explicit
Per-frame feature matrix over time 1D temporal convolution Local to long-range patterns
Player sets over time Deep Sets or Set Transformer with a TCN Order-free players, no fixed slots
Player graphs over time Graph network with LSTM or attention Interactions between players
Occupancy maps over time Spatio-temporal CNN The match as a moving field

The model discovers that a pass in flight still belongs to the passing team without anyone writing the rule, and more labelled matches make it better without redesign. That is its case.

If time gets its own stage

  • Noise dies without retraining. A weak away spike inside a home pass is suppressed by the decoder. The evidence stage is untouched.
  • Every change is visible. The raw scores stay on file, so the decoder’s edits can be listed frame by frame.
  • Each stage has its own test. Evidence is tested on fabricated scenes with a known answer. The decoder is tested on sequences.

Stage one can be a trained classifier or a set of physical rules. Stage two is a choice of its own: Seven ways to decode a match.

What breaks each in real football

  • One model: flicker. Outputs alternate through a scramble, and every flip cuts a possession in two.
  • One model: provider shift. Tracking from one vendor is not tracking from another. Noise, gaps and the meaning of “possession” all move.
  • Two machines: overconfident evidence. If stage one shouts, stage two cannot correct it. The scale of the scores matters as much as their sign.

How to judge it

Train and test on whole, separate matches. Score the raw per-frame output alone, then the same output behind a decoder, on both official metrics. If a decoder improves a multi-frame model, the model was not carrying the temporal load. If a decoder lifts the frame score and drops the spell score, it is smoothing away football.

Our read

Separating evidence from decision is the single most useful cut in possession inference. It turns one opaque model into two questions that each have a right answer. One model is the stronger choice only when labelled matches from the same kind of tracking are plentiful.

Our call

ConstraintCompetition rules

Why the engine is split in two

The fixture has no labels, and the rules forbid training or tuning on its answers. A learned end-to-end classifier had nothing legitimate to learn from, so football evidence is scored by fixed physical rules and an exact decoder writes the sequence. Seven dials remained, all of them set away from the match.

Sources

  1. Borghesi et al. (2023), ball possession with a temporal convolutional networkdoi.org
  2. Zaheer et al. (2017), Deep Setsarxiv.org
  3. PyTorch-TCNgithub.com
  4. PyTorch tutorial: sequence models with a CRF layerdocs.pytorch.org
  5. Technical report: Who has the ball?github.com