I am a novice in Bayesian networks. I have a problem which is best described (at least I think so) in the following story.
One wants to predict earthquakes. Let's say it has 5 variables, the last one being the probability of earthquake happening. Out of remaining four input variables, each of the last two variables depend on first two variables. The output depends on all of the 4 input variables.
Let's assume that one has a built a lab which can simulate earthquake perfectly. Therefore, all variables are observed (including output). After generating enough simulation data, the method has to be tested in the real environment. Now first two input variables are observed from sensors deployed on the ground, but the last two variables are not (I expect the network to infer them in a probabilistic manner, depending on the probability distributions of first two input variables).
Now my question is:
Do you model the last two input variables as hidden or observed in the model while training with simulated data? I will ultimately ask for evidence on the output variable.