Hey Mark, I see the forum is flourishing :D
You told me that Genie can Learn structure from continuous variables using PC algorithm but that it doesn't learn the joint distribution of the graph.
I'd just wanna make sure the data are not somehow discretized before the structure learning itself begins but that rather distribution parameters are learned for each variable/node from the data first, I suppose Gaussian is fit to them(?); how does it actually work?
thx
Structure learning with continuous variables
Re: Structure learning with continuous variables
The PC algorithm for continuous data uses a statistical partial correlation test to establish conditional independence. This test assumes linear dependencies among the variables and a multivariate Gaussian distribution, but it's quite robust against deviations. We are thinking about estimating the Gaussian parameters in the learned networks automatically, but this requires quite a lot of changes in GeNIe (and SMILE) which we are working on right now.frito wrote:You told me that Genie can Learn structure from continuous variables using PC algorithm but that it doesn't learn the joint distribution of the graph.
I'd just wanna make sure the data are not somehow discretized before the structure learning itself begins but that rather distribution parameters are learned for each variable/node from the data first, I suppose Gaussian is fit to them(?); how does it actually work?