1

I have two different datasets on which I've applied the same learning (K2) algorithm to learn a Bayesian network. I have the conditional probability table (CPT) of the class variable, for each of the two Bayesian networks learnt.

I want to compare the two CPTs, and want to be able to say that there is no significant difference between the distributions, something along the lines of statistical significance. I've heard that computing p-values might be a good way to quantify the difference, but I'm not sure how to implement it for this case.

What method would one recommend for this?

user48733
  • 13
  • 3

1 Answers1

1

BN's are generative models, I therefore recommend having a look here for some ideas on how you can compare two Bayesian Networks. Also this presentation has a list of BN scores that you can use to compare the two networks.

If, on the other hand, what you want is to compare the multinomial/binomial distributions between the CPT's (of the same Parent-Child) from two different networks, I would recommend looking here for an example of running and interpreting a Chi-squared independence test.

Community
  • 1
Zhubarb
  • 7,753
  • 2
  • 28
  • 44
  • The links are very helpful. But I'm still stuck on how [this](http://stats.stackexchange.com/questions/76898/test-whether-two-multinomial-samples-come-from-the-same-distribution) could be implemented to my case. The two CPTs of the class variable look something like this. Network1: P(class=0) = 0.7, P(class=1) = 0.3; Network2: P(class=0) = 0.9, P(class=1) = 0.1. How could I use a Chi-squared test on two such CPTs? Also, my interest is in saying that the distributions are the same. The link above says a large p-value may not imply that, and that it only _rejects_ the null hypothesis. – user48733 Feb 23 '15 at 10:25