[dialtone]

The cohort semantic meaning is stable, although not disclosed an ML system would learn its correlation to a given goal.

Cohort membership changes pretty frequently instead. So the system may put all people that browse mostly golf sites together in cohort 12345 that only the algorithm knows it's about golf sites, people enter and leave that cohort on a daily basis and you can only be a member of a single cohort at a time.

[dataflow]

Why would the cohort membership change frequently? Isn't it based on your browsing habits? I don't think my habits change frequently—do most people's?

Also, even if I take for granted that everyone's cohort changes daily, how does that imply anonymity? Like say my habit is that I check emails a ton on Monday, go on YouTube on Saturday, read the news on Sunday, etc... so my habits are changing daily, okay, but not weekly, right? Or maybe I do them in a different order on another week, but I'm not going to develop 1000 different habits across 1000 days, right? Shouldn't some kind of frequency analysis provide fairly consistent results?

[draz]

Cohorts cannot be too small (or they are not published), nor too big (or they are not particularly useful for capturing a particular set of behaviors/interests). The algorithm will balance these two constraints which will lead to any individuals coming in and out of particular cohorts. The semantic meaning of a cohort will likely change over time as well. For that, FLoC is proposing adding version IDs

[dataflow]

This is useful to know, but I'm confused how this addresses what I wrote in the above comment? It wasn't relying on cohort size being small or large.

[draz]

“Why would the cohort membership change frequently? Isn't it based on your browsing habits?” -> cohorts need to be rebalanced over time, so your cohort membership could change.