Sounds like a machine learning term that escaped from researchers into the wild. Machine learning people like to make up fun names for otherwise complicated and hard-to-summarize methods. See eg "BERT".

The name is actually descriptive. It is an algorithm for constructing semantically interesting cohorts of similar users, Locally on each user's machine.

It's actually a really good idea and certainly a lot more "privacy preserving" than anything that relies on sending fine-grained user data back to a central server for processing.

Of course there are problems with it, and I'm mixed as to whether it's something that non-Chrome browsers should even try to support.

The fact that all websites are included by default, and it's up to the individual website to opt out of inclusion, makes me squirm.

But the name makes sense and I think the core idea is a step in the right direction.