Automating causal discovery and inference like AutoML?

[siebert-julien]

Dear all,

(disclaimer: I am pretty new to causal inference and causal discovery, maybe what I am asking here has already some solutions implemented or it is just not possible to do)

I was wondering if there where already methods similar to AutoML for causal inference (including causal discovery, identification, estimation, refutation / evaluation). When I say like AutoML, I mean being able to do hyperparameter optimization, model selection for discovery and/or estimation, etc. I am well aware that without controlled experiment, there is no proper ground truth, but I was wondering, maybe there are some clever tricks to reduce the amount of work when it comes to do the full pipeline (starting from domain knowledge until the results of a causal inference analysis).

I ask because in most of the project I work in I often see the following pattern:

1. There are some prior domain knowledge about some causal relationships between variables, but the full causal graph (or the one that would be relevant for the problem at stake) is not known.
2. Causal structure discovery algorithms are making certain assumptions (assuming unknown confounders, linear vs. non-linear causal effects, independent Gaussian noise vs. all other kind of non Gaussian ones, etc.) and finding the "right" algorithm for the project is a nightmare (plus they have all sorts of hyperparameters to choose from, some works only for categorical/continuous/binary/... variables, etc.).
3. Once a causal graph is assumed, identification is pretty straightforward (in theory: just apply the do-calculus, assuming that the causal graph is good enough for the current problem).
4. Estimation algorithms: here again many methods with many hyperparameters.
5. Refutation (this is what I find awesome about dowhy): this is probably what matter the most in the context of a project: stakeholders want to know how much they can trust the analysis.

Best, Julien