Unsupervised ML
Tactical Clustering
Unsupervised discovery of defensive blocks, pressing schemes, and transitions — cluster patterns without a single hand-labeled event.
What it is
Clustering asks the corpus what patterns exist, instead of telling it what to find. Movement windows that land near each other in signature space are grouped into recurring types — no taxonomy, no event dictionary, no hand-labeled training set.
In internal evaluation, unsupervised clustering separated distinct movement archetypes — sprints, decelerations, turns, duels — before anyone named them. Naming is where the human comes in: clusters arrive with exemplars, and interpretation stays with the analyst.
How it works
Embeddings are clustered offline; each cluster carries exemplar windows, population statistics, and drill-down evidence. Because the representation is invariant to camera and body size, windows group by movement — not by appearance, kit, or venue. Method, corpus size, and caveats are documented in the evaluation suite.
What you get
Clustering discovers; tactical fingerprinting compares what discovery surfaces across teams and weeks. Together they replace weeks of manual video coding with an evidence-first review loop.
| Parameter | Value | Notes |
|---|---|---|
| Input | indexed movement windows | The same substrate search and profiling use. |
| Method | unsupervised clustering | Grouping in signature space; parameters documented. |
| Output | archetypes + exemplars | Each type with its most representative windows. |
| Reading | requires human interpretation | Clusters are evidence, not conclusions. |
Status & roadmap
Prototype. Archetype clustering runs in the internal workbench on a proof-of-concept corpus; the numbers and their framing live in the evaluation docs. Pressing schemes and defensive blocks are the intended reading of these clusters — surfaced as patterns, interpreted by people.