I am doing some reading about Bayesian networks and how to represent them with a DAG.
I have a question about how to initialize the distribution properties of the nodes. Say there is a Bayesian networks with both continuous and discrete data.
My first question is could we use the training data and use the center of mass and standard deviation (assuming univariate nodes) to initialize the prior distribution parameters. Is there any cons to doing this?
What about random initialization? Does this matter if the prior is very far away from the true distribution?
Are there any other strategies one could use?
