k2 ordering

frito · Post by **frito** » Mon Sep 15, 2008 2:40 am

How does K2 structure learning algorithm choose its ordering in Genie?

mark · Post by **mark** » Mon Sep 15, 2008 8:43 pm

We don't have the K2 structure learning algorithm in GeNIe (or SMILE). We do have the K2 prior over parameters that is used in Greedy Thick Thinning and Naive Bayes, but in neither case an ordering is required.

frito · Post by **frito** » Mon Sep 15, 2008 9:23 pm

mark wrote:We don't have the K2 structure learning algorithm in GeNIe (or SMILE). We do have the K2 prior over parameters that is used in Greedy Thick Thinning and Naive Bayes, but in neither case an ordering is required.

So it is not related to the Coopers K2 algorithm? Please see here http://www.cs.ubc.ca/~murphyk/Software/ ... ge.html#K2
this implementation needs a priory Ordering.

mark · Post by **mark** » Mon Sep 15, 2008 10:19 pm

They are two different things. (Although it may be that Cooper also used the K2 prior in his K2 algorithm and that's how it got its name. Not sure though.)

frito · Post by **frito** » Sun Sep 21, 2008 12:17 am

mark wrote:They are two different things. (Although it may be that Cooper also used the K2 prior in his K2 algorithm and that's how it got its name. Not sure though.)

Heres the coopers algorithm (http://citeseerx.ist.psu.edu/viewdoc/su ... .1.88.4436)

g(,) function is I think the marginal likelihood for i-th node and its parents

frito · Post by **frito** » Sun Sep 28, 2008 1:20 am

mark wrote:We don't have the K2 structure learning algorithm in GeNIe (or SMILE). We do have the K2 prior over parameters that is used in Greedy Thick Thinning and Naive Bayes, but in neither case an ordering is required.

Can you please elaborate on the K2 prior, I cant find any alternative structure learnign that would not refere to the coopers method, but his method requires an ORDERING...?

thx

mark · Post by **mark** » Mon Sep 29, 2008 8:07 pm

Look at Heckerman's tutorial on learning with Bayesian networks for approaches that don't require an ordering over the variables.

frito · Post by **frito** » Wed Oct 01, 2008 1:09 am

mark wrote:Look at Heckerman's tutorial on learning with Bayesian networks for approaches that don't require an ordering over the variables.

I am writing a thesis and I used hill climbing wiht k2 from Genie for some stuff, I need to write a little note about the methods I used so if u please can tell me more about your k2 algorithm implemented in Genie?

thx

mark · Post by **mark** » Wed Oct 01, 2008 1:44 am

The GreedyThickThinning algorithm in GeNIe is based on Heckerman's tutorial. The tutorial also discusses the BDe prior, one of the options for a prior in GeNIe, the other being the K2 prior. The K2 prior is defined in the following way. The prior distribution of every parameter associated with a node and a combination of its parents has a Dirichlet(1,...,1) distribution. In Latex notation it would be something like P(\theta_{X_i|Pa(x_i)})=Dirichlet(1,...1,). The number of ones depends on the number of states in the node. Let me know if you still have questions.

frito · Post by **frito** » Wed Oct 01, 2008 2:15 am

mark wrote:The GreedyThickThinning algorithm in GeNIe is based on Heckerman's tutorial. The tutorial also discusses the BDe prior, one of the options for a prior in GeNIe, the other being the K2 prior. The K2 prior is defined in the following way. The prior distribution of every parameter associated with a node and a combination of its parents has a Dirichlet(1,...,1) distribution. In Latex notation it would be something like P(\theta_{X_i|Pa(x_i)})=Dirichlet(1,...1,). The number of ones depends on the number of states in the node. Let me know if you still have questions.

many thanks Mark!