Is there confidence interval for the parameter estimation?

[BayesFusionUser123]

Dear Community,

I am following the PDF version of the tutorial for SMILE wrappers, specifically for python language.

I am try SMILE package on our real-world dataset.

Since the real world dataset is too big. I have to use "pysmile.BayesianAlgorithmType.EPIS_SAMPLING" to produce inference result. I found that every time I print out the posterior of the target variable, it changes. Even though I set a specific seed, the result has variability each time I run.

My use case for SMILE is to do effect estimation on a parameterized Bayesian network. Suppose there is a network A -> T. A has two values: A_0 and A_1. I firstly use set_control_value function to set A as A_0, print out posterior of T. Then clear the control value, then set control value A as A_1. Then print out the posterior of T. Then I calculate the diff between the posterior of T.

Currently, I uses 10 seed, print posterior 10 times and then calculate mean and standard deviation as my final result.

I have the following questions:

(1) Why even though I set a specific seed, each time I print the posterior of target variable, it still has variability?

(2) During the EM algorithm to learn parameter of the network, is there confidence interval for parameter estimation? Can the inference step take the confidence of the parameter estimation into account?

For instance, when we estimate P(A=1), if we have 10 samples, 2 positive, we will have P(A=1)=0.2. If we have 100 samples, and 20 positives, we will also have P(A=1)=0.2, but the confidence interval for the 0.2 is much tighter for the larger sample size.

Thank you very much for your answer.

[shooltz[BayesFusion]]

(1) Why even though I set a specific seed, each time I print the posterior of target variable, it still has variability?

This should not happen, the pseudorandom generator is seeded from the value specified in network properties. If you can share your model, we can investigate this.

(2) During the EM algorithm to learn parameter of the network, is there confidence interval for parameter estimation? Can the inference step take the confidence of the parameter estimation into account?

No, there are no confidence intervals calculated during the EM parameter learning.