see also:
consider some community that contains subcommunities, such that membership of a community is part of that members personal identity.
consider some specific subcommunity such that is comprised of all members of A that share some common interest. The characterization of the associated “interest” ultimately defines a neighborhood around it. As the interest gets narrower, the radius of that neighborhood gets smaller.
Personalization is inherently a process of allocating people into narrow interest communities.
This necessarily pushes fringe elements together, and binds them as the associated interest gets more narrowly defined and crystallized
now, consider the directed community membership probability thing I came up with for bimodal projections.
From the perspective of members, they still identify as members of , even if their values are fringe minority within that membrship. for all x in a’, x is also in A.
But from the perspective of A, a’ is fringe. P(x in a’ | x in A) is small by construction, and as A grows (or conversely, as the interest that characterizes a’ shrinks), the relative asymmetry of the relationship gets wider.
Large communities naturally subdivide into smaller communities like this. I’d posit that it’s also in their best interest to do so: factorizing into a hierarchy of subgroups is good for resilience and adaptability. But this also is clearly a process that can naturally drive ideological radicalization.
It could be beneficial to factorize in this way in the presence of a pruning function. This is why volunteer content moderation isn’t enough: community platforms also need to curate which communities they allow on the platform.