I haven’t read enough of Causality, but I think I get how to find a causal model from the examples above.
Basically, a model selection problem? P(Model|Data) = P(Data|Model)P(Model)/P(Data) ~ P(Data|Model)P(Model)?
Is P(Model) done in some objective sense, or is that left to the prior of the modeler? Or some combination of contextually objective and standard causal modeling priors (direction of time, locality, etc.)?
Any good powerpoint summary of Pearl’s methods out there?
P(Model) is usually related to the dimension of the model (number of parameters). The more parameters, the less likely the model (a form of the razor we all know and love).
There are other ways of learning causal structure, based on ruling out graphs not consistent with constraints found in the data. These do not rely on priors, but have their own problems.
I haven’t read enough of Causality, but I think I get how to find a causal model from the examples above.
Basically, a model selection problem? P(Model|Data) = P(Data|Model)P(Model)/P(Data) ~ P(Data|Model)P(Model)?
Is P(Model) done in some objective sense, or is that left to the prior of the modeler? Or some combination of contextually objective and standard causal modeling priors (direction of time, locality, etc.)?
Any good powerpoint summary of Pearl’s methods out there?
Hi,
P(Model) is usually related to the dimension of the model (number of parameters). The more parameters, the less likely the model (a form of the razor we all know and love).
See these:
http://en.wikipedia.org/wiki/Bayesian_information_criterion http://en.wikipedia.org/wiki/Akaike_information_criterion
There are other ways of learning causal structure, based on ruling out graphs not consistent with constraints found in the data. These do not rely on priors, but have their own problems.