What does this prove? If you have lots of data and dimensions, I bet you could just as likely find distributions that are roughly 50/50, 60/40, 90/10, 100/0 if you looked for them.

Agreed. It’s not necessarily 80/20, it’s just that power laws show up a LOT. 90/10, 99/1, etc.

So you massage your features until they can produce a 20-80% split?

Very scientific.

The 50/50 principle always shows up in my histogram with interval of two.