search

py-why / dodiscover

[Experimental] Global causal discovery algorithms
https://www.pywhy.org/dodiscover/
MIT License
82 stars 17 forks source link

Repeated separating set testing as a subprocedure #119

Open adam2392 opened 1 year ago

adam2392 commented 1 year ago

Is your feature request related to a problem? Please describe. Faithfulness is commonly violated. One aspect of the problem is the (incorrect) orientation of colliders, which then lead to (incorrect) orientation of downstream edges.

Describe the solution you'd like https://proceedings.neurips.cc/paper/2019/file/e6872f5bbe75073f8c7cfb93de7f6f3a-Paper.pdf introduces a simple procedure for iteratively running PC/FCI algorithms.

We should implement a similar abstraction, but make it optional for users since it essentially results in a drastic increase in computation.

adam2392 commented 1 year ago

Supposedly this has a bug in the procedure, as raised by Adele Ribeiro via some counter-examples.