Causal Discovery Algorithms

23 Sep 2021 09:45

Yet Another Inadequate Placeholder, because I should really split Graphical Causal Models into (at least) two notebooks, this being one.

Things I need to think/learn more about:

  1. How important is faithfulness? How unreasonable is it, and what strictly weaker assumptions will do (most of) the same work?
  2. Confidence sets for the causal graph. The simplest approach is to go over every possible graph, test it, and return the ones which pass the test. But this is computationally infeasible (the number of graphs grows crazily fast with the number of variables) and hard to interpret. Are there more comprehensible representations possible, and/or ones which would make this more computationally tractable?