> Doesn't that feel like a repeat with whats happening in social media and news media these days? Or is it just different things.
Maybe. It can feel like this, but applying mathematical models to social phenomena is tricky. You have to map formal parameters to fuzzy, poorly understood factors. How do you define, precisely, what is a "channel" on a social network? What is "signal", and what is "noise"? People can and do prove anything by using slightly different (or inconsistent) definitions here, getting the numbers to line up just like they want them to.
My understanding is that a better way to approach such mapping would be to shove in probability distributions in place of hard-to-map exact parameters - abstracting away choices and measurements lets you see the wider context here. This pushes the problem to defining the appropriate distributions, but I think that's more tamper-proof. Unfortunately, the results may come out next to useless - e.g. probability distributions so wide you could sail a carrier strike group through them.
I'd love to know what's considered the correct, robust approach to such problems.
I think OP has a good point. Although language if fuzzy, so are "binary" signals once you drill down to the physical layer. The fuzziness of language goes away if you allow yourself to assume an idealized language embedding model. Similar ideas are in an epsilon ball around the embedding of some concrete linguistic instantiation of the idea.
An example of a binary signal: "It is unambiguously true that a lab leak did not occur in Wuhan." The spread of this binary signal could be measured using the technique described above. A BERT embedding may suffice.
The channels are just tweets, DMs, and Vox articles.
Maybe. It can feel like this, but applying mathematical models to social phenomena is tricky. You have to map formal parameters to fuzzy, poorly understood factors. How do you define, precisely, what is a "channel" on a social network? What is "signal", and what is "noise"? People can and do prove anything by using slightly different (or inconsistent) definitions here, getting the numbers to line up just like they want them to.
My understanding is that a better way to approach such mapping would be to shove in probability distributions in place of hard-to-map exact parameters - abstracting away choices and measurements lets you see the wider context here. This pushes the problem to defining the appropriate distributions, but I think that's more tamper-proof. Unfortunately, the results may come out next to useless - e.g. probability distributions so wide you could sail a carrier strike group through them.
I'd love to know what's considered the correct, robust approach to such problems.