0
$\begingroup$

Suppose I have a Bayesian network that can be factorized like this:

P(A,B,C,D)=P(A)*P(B)*P(C|A,B)*P(D|C)

Each of the variable is a binary and I've got all the tables of conditional probabilistic distribution.

I then simulate from this network and got a synthetic data-set:

    A  B  C  D
    1  0  0  1
    0  1  0  1
    1  0  1  1
    etc. 

I then estimate the parameters (the probabilities) with MLE and Bayesian estimator from the synthetic data and would like to compare these 2 learners by their KL-distances against the true distribution at various sample sizes.

However, I'm not sure how to do it. The book I read about the relative entropy of Bayesian network just has notations in it and it's very hard to follow.

Thanks!

$\endgroup$
1
$\begingroup$

Each of your models is a joint distribution for the random variables $A, B, C, D$. This means you can calculate the probability according to each model that the random variables take particular values: $p(A=a, B=b, C=c, D=d)$. I'll use the shorthand $p(a, b, c, d)$ to mean the same thing (uppercase letters are random variables, lowercase are particular values).

Say the the true distribution is $p$ and the model is $q$.

In your case, the formula for KL divergence is:

$$ D_{KL}(p \parallel q) = \sum_{a, b, c, d \in \{0, 1\}} p(a, b, c, d) \log \frac{p(a, b, c, d)}{q(a, b, c, d)} $$

The sum is taken over all possible values of $A, B, C, D$. You have 4 variables with 2 possible values each, so there are 16 possible joint configurations: $\{A=0, B=0, C=0, D=0\}$, $\{A=0, B=0, C=0, D=1\}$, etc. Loop over the 16 possible configurations. Determine the probability of each configuration according to each model. Plug those values into the expression above, taking the sum over all configurations.

$\endgroup$
  • $\begingroup$ Thanks, but what if there are more nodes that makes it hard to enumerate the probability of each of its configuration? I've seen a formula in Koller and Friedman that I can somehow use the conditional probability distribution table without calculating the full joint, but that's where I got stuck actually. @user20160 $\endgroup$ – mackbox Aug 2 '16 at 23:20

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.