Once each stroke could be represented in the compressed form, it was possible to compare them.
Two putting strokes might look similar at impact but have slight differences in the tempo, symmetry, micro-movement, etc. Although these are hard to isolate and analyze with traditional methods, they show up naturally by comparing learned representations.
Instead of asking “Was this putt successful?”, I could ask “Which strokes’ motions are similar?”
By grouping strokes based on their internal representations, patterns emerged. Some clusters showed highly repeatable, smooth trajectories while others contained strokes with the same overall shape but different timing. Some groups captured motions that looked controlled but relied on late corrections, while others reflected more uniform movement throughout the stroke.
What mattered was that these groupings were not imposed, but rather emerged directly from the data. In many cases, strokes from the same golfer landed in multiple clusters, highlighting variability that outcome-based analysis often ignores.
This step was critical because it decoupled motion from results. A cluster wasn’t “good” or “bad”, it was simply a recurring way of moving. Performance could be examined later, but motion came first.
That separation turned out to be one of the most important conceptual shifts in the project.