Unsupervised clustering

Archetypes

We stripped the names off all 15 profiles and asked a k-means model to group them by structure alone — industry, how the wealth arose, education, age, scale. It found 2. Every claim below prints the counts behind it, and the sample is small enough that you should read this as a sketch rather than a result.

n = 15k = 2silhouette 0.31deterministic (seed 42)
Groups found
2
Best silhouette across k=2…5, rejecting any k that isolated a single person.
Separation
0.31
Mean silhouette. Real but overlapping — well below the 0.5+ of cleanly separable groups.
Features used
12
Wealth (logged), age, Forbes self-made score, education level, gender, company count, industry.
Largest group
10 people
Self-made builders

Choosing the number of groups

Shown rather than asserted. Silhouette keeps climbing with k because more clusters always fit a small sample better — at this size it peaks by isolating individuals, which is exactly why the minimum-size rule matters.

kSilhouetteGroup sizesVerdict
20.3115 · 10chosen
30.0931 · 3 · 11rejected — isolates an individual
40.3541 · 1 · 4 · 9rejected — isolates an individual
50.4501 · 1 · 4 · 8 · 1rejected — isolates an individual

Note what this table admits: the highest raw score belongs to a k we refused. Reporting only the chosen row would have made the model look better than it is.

The groups

Group names are OUR shorthand for what each centroid contains — a description we wrote, never something the model discovered.

Self-made builders

10 people · median $213B · median age 53

02

Over-represented here

Built from little (Forbes 8–10)
9/10 vs 9/151.5×
Technology
8/10 vs 8/151.5×
University dropout
4/10 vs 4/151.5×
Postgraduate degree
3/10 vs 3/151.5×
Self-made
10/10 vs 11/151.4×

Inherited retail dynasties

5 people · median $126B

01

Over-represented here

Fashion & Luxury
5/5 vs 5/153.0×
Inherited
4/5 vs 4/153.0×
Pure inheritance (Forbes 1–3)
2/5 vs 2/153.0×

Lift is the share inside a group divided by the share across all 15. A lift of 3× means the attribute is three times as common here as across the whole set. Attributes held by only one person are excluded — one person is an anecdote, not a driver.

How this was computed

  • Net worth is logged. The top of this set is many times the bottom. On a raw scale every boundary would be “is this the richest person”, which is one outlier rather than a pattern.
  • Every column is standardised. Otherwise age would dominate distance purely because its numbers are larger than a 0/1 industry flag.
  • Country was measured, then dropped. Including nationality lowered separation and split coherent groups along national lines. Where someone is from is already its own chart.
  • The seed is fixed. k-means normally starts from random centroids, so the page would show different groups on each reload. A reader cannot check a claim that changes while they look at it.

What this cannot tell you

15 people is not enough for real clustering. At this size the algorithm is describing a handful of obvious contrasts — mostly self-made technology against inherited consumer brands — which you could see by reading the table. Treat it as a structured restatement of the dataset, not as a discovery.

It cannot predict who becomes wealthy. Everyone here was selected because they are already among the richest people alive. There is no comparison group of people with identical attributes who did not get here, and there are many millions of them.

A silhouette of 0.31 means the groups overlap substantially. Several people sit near the boundary and could reasonably belong to either side.

Forbes’ self-made score is an editorial judgment, not a measurement. It is a strong input here, so its subjectivity propagates into every group.

← Pattern analysis·About the data →