Using DAGMA for inference and mixed-type data

[JulianMeigen]

Hi Kevin,
thanks for releasing DAGMA.
I am currently exploring ways to improve a tool/project that uses conditional-Gaussian Bayesian networks to model dependency structures in datasets with mixed data types (implemented in R using bnlearn). I am evaluating differentiable structure learning approaches such as DAGMA/NOTEARS as potential alternatives.

Since I am still getting familiar with these methods, I had two questions:

1. Inference / sampling:
   Is it possible to perform Bayesian network inference (e.g. sampling) using the DAG resulting from DAGMA, or should DAGMA mainly be understood as a structure learning method that requires an additional parameter-learning step to obtain a full generative model?

2. Mixed data types:
   In my setting, the data contains both categorical and continuous variables. For parameter learning in conditional-Gaussian BNs, the structure must be restricted such that discrete nodes only have discrete parents. Is there a principled or practical way to impose such constraints when using DAGMA, or a recommended workflow for handling mixed data?

Any guidance, references, or suggested workflows would be greatly appreciated.

Best regards
Julian