I have yet to wrap my head around why do-Calculus and Pearl's causal inference is in any way more useful than the many statistical methodologies for causal inference that have existed for a century. Every time I've tried to dive into Pearl's justifications for what makes his methodology superior, he mostly spends time attacking the character of long dead statisticians (and some living ones), and bemoaning how awful the Rubin causal paradigm is, without giving any actual, verifiable specifics. I also have yet to see just about anyone using do-Calculus in a practical way.

If it really is all that, I'm interested and please explain it to me. Because up until now it seems like its basically someone who reinvented the wheel, and has formed a cult following around it by saying "but MY wheel has no death crystals!"

I see the do-calculus as being causality as applied to a graph. It then gives you a bunch of correlations to test.

The other example of "things applied to graphs" is graphical models aka Bayesian networks => Bayesian statistics.

Causal inference is only helpful is fields where an intervention is needed (like social science, medicine, etc.), which limits its applicability somewhat. Sure, ML is only correlation, but for most problems correlation is all you need.

It's quite useful in business functions such as marketing

I think that in reality it is helpful in most situations, as you are most likely doing the analysis to decide on actions to do, which are interventions and thus assume a causal graph. It’s just that, normally, the impact of misidentifying a correlation as a cause is cheap, and corrected soon enough, so even if it is helpful, it is not essential, and it might not even justify the extra effort.

Any ML model that affect people can be analyzed from a social science point of view. Maybe causal inference could be used for model fairness/bias

Has there been any good work applying causal inference explicitly to common NLP tasks like question-answering? Or to game-play? The lack of common-sense in models like GPT-3 would seem to indicate a huge opportunity, perhaps in tandem with COMET[0] or some other graph-based approach. [0] https://arxiv.org/abs/1906.05317

inference is so fundamental to ML there should be classes just taught from the perspective - currently you'll encounter it in a graphical models course and when discussing networks of course, but again, viewing ML from the lens of inference like pearl does should really be it's own required ML course