Which parameters influence Bayesian score ?
-
codimension.one
- Posts: 2
- Joined: Wed Nov 13, 2013 7:19 am
Which parameters influence Bayesian score ?
I am trying to compare several models derived via the Bayesian search procedure in GeNie, and I am not sure which search parameter values will influence the Bayesian score of the resulting structure. For example, if I obtain a particular graphical structure with one particular choice of link probability and prior link probability values, then obtaining that same graphical structure with a different set of link probabilities (and/or prior link probabilities) will result in that same graphical structure being assigned a different Bayesian score. My question is which other search parameters will cause identical structures to be assigned different scores ? (For instance, will the 'sample size' parameter or the 'discrete threshold' parameter do this ?)
Re: Which parameters influence Bayesian score ?
Hi,
These are the parameters that are used for the Bayesian Search algorithm, and I've listed if they influence the score or not (i.e. if the same structure obtained under different parameter settings would result in a different score)
Max Parent Count - Limits parents of a node, has no effect on the score, i.e. same network found using different values results in same score.
Iterations - Number of Random restarts, has no effect on the score
Equivalent Sample Size - influences the score, part of the BDeu calculation, it's sort of represents an inertia of the current parameters when introducing new data.
Seed - Seed used for generating random starting point, has no effect on the score
Link probability - parameter is used when generating the random starting network, has no effect on the score.
Prior Link Probability - influences the score, it's a prior over all edges. (works in the following way: log Posterior score = log marginal likelihood (i.e. the BDeu) + |parents|*log(pll) + (|nodes|-|parents|-1)*log(1-pll))
Best,
Martijn
These are the parameters that are used for the Bayesian Search algorithm, and I've listed if they influence the score or not (i.e. if the same structure obtained under different parameter settings would result in a different score)
Max Parent Count - Limits parents of a node, has no effect on the score, i.e. same network found using different values results in same score.
Iterations - Number of Random restarts, has no effect on the score
Equivalent Sample Size - influences the score, part of the BDeu calculation, it's sort of represents an inertia of the current parameters when introducing new data.
Seed - Seed used for generating random starting point, has no effect on the score
Link probability - parameter is used when generating the random starting network, has no effect on the score.
Prior Link Probability - influences the score, it's a prior over all edges. (works in the following way: log Posterior score = log marginal likelihood (i.e. the BDeu) + |parents|*log(pll) + (|nodes|-|parents|-1)*log(1-pll))
Best,
Martijn