At a glance
- 6 standardised features over 13 weeks
- k chosen by silhouette score (2–5)
- Bootstrap Jaccard stability over 100 resamples: under 0.6 unstable, 0.6–0.75 a pattern, 0.75–0.85 stable, over 0.85 highly stable
- Segments named by what makes them most distinctive
How it works
- Features are standardised (z-scores) so no single unit dominates.
- K-means runs for k = 2…5 and the best silhouette is kept.
- PCA projects the 6 dimensions onto 2 for the chart only.
- The R validation script re-checks with the gap statistic and MANOVA.
How to read it
- Treat unstable segments as hypotheses. Shelly says so rather than overselling.
- Waste-exposed perishables → tighter ordering and earlier markdowns. Traffic drivers → never out of stock.
Limits & safeguards
- On demo data the silhouette is modest (~0.28) and the gap statistic prefers 2 clusters. That's reported honestly.
Ask Shelly
Tell me about the segments
Type or say these in the chat on the main page.