reduce dimensionality of floc set with a bloom filter to reduce privacy budget impact about floc HOT 2 OPEN

Here is a document on the theoretical properties of a decaying bloom filter

https://openproceedings.org/2013/conf/edbt/DautrichR13a.pdf

from floc.

There are some plausible ways of building flocks that start with page taxonomies, as you suggest. It seems like it would require on-device ML to figure out the taxonomy of each page (in many languages!) — not impossible, but obviously additional work.

But I don't think the one-sided false-positive properties of a Bloom filter are particularly helpful here. For example, this would mean a person's flock would give out the definite signal "I have not visited a page about Topic T in the last D days." And if a person visits the same site regularly and that site sees their flock change, it's very likely that the change will indicate exactly what new topic of page they visited in between.

We plan to experiment with a range of approaches to clustering.

from floc.