Grouping Players Who Move the Same Way

Opening Scene

After watching the pickup scrimmage from Article 2 for twenty minutes, an assistant coach starts sketching rough groups on a notepad — not positions, nobody’s assigned those, just clusters of players who seem to move alike. One group sticks near the basket and rarely sprints. Another covers huge amounts of ground and rarely stops moving. A third hovers at mid-range, mixing both. No rule was written down for this; the assistant just kept comparing players to each other and adjusting the groups until they felt stable.

That instinctive process — comparing everyone to everyone, adjusting groups until they settle — is almost exactly what a clustering algorithm does mechanically, at a scale no assistant could match by eye.

In Plain English

Clustering groups data points together based on how similar they are to each other, with no labels telling the algorithm what the “right” groups are. A common approach, k-means, picks a number of groups, assigns every point to whichever group center it’s closest to, recalculates the centers based on the new groupings, and repeats until the groups stop changing. The output is a set of groups the data itself suggested — not ones anyone defined in advance.

The Old Way

Before clustering had a formal name, people did rough versions of it constantly:

A store manager eyeballing which customers tend to buy similar things and mentally bucketing them into rough segments.
A librarian noticing which patrons check out similar books and grouping them loosely by reading taste.
A teacher noticing which students struggle with similar material and forming informal study groups around it.

This worked at small scale because a person could hold the whole group in their head. It broke down completely the moment “the group” became too large to observe directly — exactly the gap clustering algorithms exist to close.

What’s Changing (and Why AI Is the Reason)

Clustering can now run on datasets with millions of points and dozens of dimensions at once, far beyond what an assistant sketching on a notepad could ever hold in view — the same instinct, applied at a scale that used to be impossible.
Choosing how many groups actually exist is increasingly assisted rather than guessed. Deciding “how many clusters” used to be mostly trial and error; tooling can now suggest a reasonable number based on how the data actually separates.
Clustering results are increasingly fed straight into other AI systems as a preprocessing step, not just reviewed by a human as a standalone report — a theme this series returns to directly in Article 19.

The Metaphor, Fully Extended

Pickup Scrimmage	Clustering Concept
Players grouped by how they move	Clusters — groups formed by similarity
The rough “center” of a group’s typical movement	A cluster centroid
A player who clearly belongs near the basket group	A point strongly assigned to its cluster
A player who could plausibly fit two different groups	A borderline or ambiguous cluster assignment
The assistant redrawing group lines after watching longer	Reassigning points as cluster centers update
Deciding to use three groups instead of two	Choosing the number of clusters, k

For Beginners: What to Actually Do

Try clustering on a dataset you already understand reasonably well first — it’s far easier to judge whether the groups make sense when you already know something about the data.
Don’t treat the number of clusters as fixed; try a few different counts and see which produces groupings that actually make intuitive sense.
Look at a handful of actual examples inside each cluster, not just the summary statistics — a cluster that looks clean in aggregate can still contain a strange mix on closer inspection.

For Practitioners and Leaders: The Deeper Layer

Clustering is sensitive to how the input features are scaled — a variable measured in the thousands can quietly dominate the grouping over one measured in single digits unless the data is prepared carefully.
Present clustering output as a hypothesis for a business team to react to, not a finished segmentation — the algorithm found statistical structure, and a human still needs to confirm it maps to something real and actionable.
Revisit clusters periodically; the natural groupings in a dataset shift as behavior shifts, and a clustering done a year ago may no longer reflect reality.

Quick Recap

Clustering groups data by similarity alone, with no labels defining the “right” groups in advance.
It mirrors an old human instinct — informally bucketing things by resemblance — done mechanically and at much larger scale.
Choosing the right number of groups is part science, part judgment, and increasingly assisted by tooling.
Clustering output is a starting hypothesis, not a finished answer — it still needs human or business validation.

Where This Fits in the Series

Article 4 covered the two shapes a supervised label can take; this article returned to unsupervised territory with the most common way to find groups in unlabeled data. Article 6 looks at a related but different question — not how players group, but how many dimensions actually matter in describing them at all.