[
Date Prev][
Date Next][
Thread Prev][
Thread Next][
Date Index][
Thread Index]
[
List Home]
Re: [recommenders-dev] recommenders-dev Digest, Vol 57, Issue 4
|
Hi Javier,
First of all, I would check whether all the parents really need to be modeled as parents. If (some of) the parents are actually conditionally independent given the child, you can get away with reversing the parent-child relationship.
Also I would look into noisy-or and noisy-and nodes. Both are not implemented in Jayes, but could allow to reduce the free parameters of the model.
Regards, Michael
Am 28.10.2015 16:41 schrieb "Javier Martínez Torres" <
javi.martineztorres@xxxxxxxxx>:
Thank you Michael,
Currently the Bayes Network works like a dream!! However to apply this technology in a real scenario the limitation will surely come from LCPT (Local Conditional Probability Table) size. If a node has 32 parents, the LCPT will have 2^33 double numbers (out of memory)!! To deal with this limitation I've been researching on "Divorcing techniques". Probably most of you had been involved in dealing with combinatorial explosion. My questions are:
1.- What is your suggestion to overcome the out of memory trouble?
2.- If I would decide to implement divorcing techniques, What is the algorithm to introduce "ghost nodes" without interfering the result in terms of likelihood?
Thank you in advance,
_______________________________________________
recommenders-dev mailing list
recommenders-dev@xxxxxxxxxxx
To change your delivery options, retrieve your password, or unsubscribe from this list, visit
https://dev.eclipse.org/mailman/listinfo/recommenders-dev