04Solutions family

Classifiers and decoders

Models that turn evidence into probabilities, and decoders that turn probabilities into possessions. The family our evidence decoder belongs to.

Scores for three states on every frame, and the one best path through them.
  • 11entries
  • 1fielded: Viterbi decoding
  • 4need no labels

One picture: two jobs

Evidence is sparse and noisy, and one lane goes dark when the ball is untracked. One exact decode turns all of it into a single coherent sequence.

Two jobs live in this family, and keeping them apart is the whole idea. Scoring turns each frame’s evidence into support for home, away or dead. Decoding turns those per-frame scores into a sequence that looks like football: possessions that last, changes that happen at a touch.

Scoring can be learned, with trees, convolutions or graph networks, or specified by hand. Decoding is exact mathematics either way.

This family is the last link of the chain. Everything upstream produces evidence. Here it becomes an answer.

How to choose inside it

Entry Labels Ball and identities Compute
Viterbi decoding None Neither. A hidden ball is a frame without evidence Seconds on a CPU
Hidden Markov models Optional Neither Light
Hidden semi-Markov models For durations, or assumptions Neither Grows with the longest spell
Linear-chain CRF Labelled sequences Neither Light
Gradient-boosted trees Frame labels A hidden ball is a value. No identities CPU
Set encoders Frame labels No identities, by design GPU to train
Temporal convolutional networks Frame labels, many Set by the features fed in GPU to train
Temporal action segmentation Frame labels, many Set by the features fed in GPU to train
Graph network toolkits Frame labels, many Ball node needs a fallback. Temporal layers want identities GPU to train
Online change-points None One signal Light
Learned imputation Complete sequences to learn from It fills the ball in GPU to train

Labels drive the choice. With none, the decoders are the family: Viterbi first, a semi-Markov model if spell lengths are the problem. With labels, score frames with boosted trees before anything deeper, and keep a decoder behind whatever scores them.

The family by its appetite for labels. Decoders on the top row, scorers below. Everything inside the lime zone runs with no labels at all, and the pair we fielded sits there.

What this family cannot do

  • A decoder cannot repair biased evidence. It finds the best sequence for the scores it is given.
  • A classifier cannot be trusted without labels from the target feed. It learns one provider’s possession.
  • Frame accuracy does not buy coherent possessions. A model right on most frames can still split every possession in three.

Where our decoder stands

Here. It is an exact Viterbi decode per half over H, A and D, with hand-specified football evidence and a cost for every change of state. Nothing is fitted to the match. Why this decoder and not a learned or duration-aware one is argued in Seven ways to decode a match, and the whole design is in The Shrewdies decoder.

In this family11

  • The exact best path through per-frame evidence. Three states, a cost for every switch, and a guarantee that no better sequence exists. Our evidence decoder is built on it.

    • Sequence
    • Exact
    • No training
    • Fielded
  • The generative sequence model: hidden states, noisy observations, exact inference. Forward-backward gives per-frame posteriors; Viterbi gives the single best path.

    • Sequence
    • Classical
    • Labels optional
    • BSD-3-Clause
  • An HMM in which each state chooses how long to last. The natural model for spells, and the expensive one.

    • Sequence
    • Classical
    • Duration-aware
    • GPL-3.0
  • Learned transition scores between adjacent labels, trained jointly with the network beneath and decoded exactly with Viterbi.

    • Sequence
    • Structured output
    • Needs labels
    • MIT
  • The strongest general-purpose learners for tabular data. Give each frame a row of distances, speeds and visibility flags and they return H/A/D probabilities.

    • Feature table
    • Learned
    • Needs labels
    • MIT / Apache-2.0
  • Networks whose output cannot depend on the order of their inputs. The correct way to feed anonymous players to a model.

    • Tracking
    • Learned
    • Order-invariant
    • Needs labels
  • Dilated one-dimensional convolutions that read seconds of context for every frame. The standard learned sequence layer for dense per-frame labels.

    • Sequence
    • Learned
    • Needs labels
    • MIT
  • Models built to label every frame of a long sequence without shattering it. Computer vision met the over-segmentation problem first, and its fixes transfer.

    • Sequence
    • Learned
    • Needs labels
    • Research code
  • 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
  • A running posterior over how long the current regime has lasted. When it collapses to zero, something changed.

    • Time series
    • Classical
    • Online
    • No training
  • Neural models that fill gaps in multivariate time series. Strong on sensor dropouts, risky on a ball that vanishes while the game goes on.

    • Time series
    • Learned
    • Missing data
    • MIT / BSD-3-Clause