Running Bayesian Network with 10000 nodes

[WeiHsinGu]

Hi,

I am running a Bayesian Network with 10000 nodes using Csmile library. The network has a structure which is almost a tree but has several branches (<10) converging to a bottom node.

My question is: in this situation, what update belief algorithm can lead to the best performance? Is there any tips on running such a network?

Thanks.

[shooltz[BayesFusion]]

If you can run the inference with the default exact algorithm (i.e. the DSL_network::UpdateBeliefs returns DSL_OK), then it's probably your best bet performance-wise. It may be possible to improve the speed when you're only interested in the posteriors for the subset of the network nodes, mark them as targets with DSL_network::SetTarget. The clustering algorithm can take advantage of this and reduce its workload, but in general case the improvement will depend on the structure of the network and evidence set.