1/4 Broadly, because estimating a cond. prob. P(Y | X) is easier than a joint one, P(Y, X, Z), when X & Z are high-dimensional. Indeed, this observation led to (non-causal) graphical models in the 1980s, including Bayesian nets (ie, non-causal DAGs), Markov random fields & so on.
https://twitter.com/deaneckles/status/1355624929920741378
2/4 The goal in this journey was to find parsimonious encoding of distributions that may lead to efficient learning-inference (e.g., belief propagation). There is the whole UAI community, which is now permeated throughout other ML conferences, focused on this class of algorithms.
3/4 This was insufficient for causal reasoning, as noted in the early 1990's & discussed in Sec. 1.4 here causalai.net/r60.pdf (e.g., see ex. 10). It has been a quite traumatic transition, as noted by @yudapearl , see quote found in p. 30: 
4/4 Why are BNs not enough? What is missing? Additional assumptions of modularity & better semantics for the arrows in the model are needed, which has been done by Pearl (Causality, p. 23), and refined in causalai.net/r384-lnai.pdf , and generalized R-60 (Def. 12/16).
• • •
Missing some Tweet in this thread? You can try to
force a refresh
