To answer the original question by Professor Neapolitan and the follow up by Professor Pearl, I came across a variation of the Noisy OR model - the Recursive Noisy OR Rule - A rule for estimating complex probabilistic interactions published in IEEE Transactions on Systems, Man and Cybernetics in 2004 by Lemmer and Gossink. Using this rule, the authors show how a complete CPT can be computed (in sparse data situations) as well as how expert opinion can be incorporated. Having found, no other reference of this rule, I conducted an empirical study of this rule and compared it with the Noisy OR model in the domain of childhood asthma using data from our EMR. I found no statistically significant differences in performance of a belief network using the parameters computed using Noisy OR Vs RNOR, however RNOR did provide a way to compute parameters when no data were available. My results are published in - An Empirical Validation of Recursive Noisy OR Rule for Asthma Prediction in the AMIA 2010 symposium proceedings. Therefore, If I understand the problem correctly, I would suggest RNOR as a solution, not necessarily better than Noisy Or.
Best Regards, Vibha Anand From: uai-boun...@engr.orst.edu [mailto:uai-boun...@engr.orst.edu] On Behalf Of Judea Pearl Sent: Saturday, February 12, 2011 7:57 PM To: uai@engr.orst.edu; Rich Neapolitan Subject: Re: [UAI] Learning Parameters for the Noisy-OR model Rich, Why would it be different from logistic regression, for which there is a volume of statistical literature.?. (If I take logP(x=0|pa(x)), I get a linear expression in the parameters, the rest should fall in place) ==Judea ----- Original Message ----- From: Rich Neapolitan<mailto:re-neapoli...@neiu.edu> To: uai@engr.orst.edu<mailto:uai@engr.orst.edu> Sent: Friday, February 11, 2011 8:32 AM Subject: [UAI] Learning Parameters for the Noisy-OR model Once again, I am going against the grain and submitting a post that is not a job ad or a conference announcement. I hope no one takes offense. My question concerns the noisy-OR model. The traditional way to assess a parameter value for a given cause is to use the data items that only have that cause present. However, if there are many causes and limited data, there will be few such data items. I want an approximation method that deals with this problem. A quick Google search did not reveal any previous work in this area. I have a few ideas, but I thought I would first ask if anyone knows of anything that has already been done in this area. Best regards, Rich Rich Neapolitan Professor and Chair of Computer Science Northeastern Illinois University 5500 N. St. Louis Chicago, Il 60625 ________________________________ _______________________________________________ uai mailing list uai@ENGR.ORST.EDU https://secure.engr.oregonstate.edu/mailman/listinfo/uai
_______________________________________________ uai mailing list uai@ENGR.ORST.EDU https://secure.engr.oregonstate.edu/mailman/listinfo/uai
