Hello,
I want to construct a Bayesian network given the data. When the data is complete I am able to do it using an R package DAKS. But when the data is missing this package fails. Does anyone aware of this kind of software or package which can help me to learn / draw the Bayesian network with data containing missing values. It need not be R, any language is fine as long as it is giving me the results.
Thanking you,
Irshad
What about using a Neural Network !
Its generalization property helps a lot when there is missing data...
Dear Irshad,
For Bayesian approach you can use Winbugs software which is specified only for Bayesian approach analysis.
You can find this software in here,
http://www.mrc-bsu.cam.ac.uk/software/bugs/the-bugs-project-winbugs/
GeNIe (http://genie.sis.pitt.edu/) has learning but not structural learning with missing data. You will love the user interface. I suggest the approach proposed by Paul above. Once you have learned the structure, you can learn the parameters from the original data set (with missing values). The EM algorithm implemented in GeNIe has no problems with it. I hope this helps.
Marek
thank you for the replies guys.
@Paul, my data is dichotomous so that discards the possibility of replacing the the missing values with average. And the amount of missing data is considerable so I can not even delete it.
@Jorge Lopez-Cifre, thank you for the reply.
Can you please provide me some material on that.
@Abdur, I doubt can you please give me some documentation of winBug. The link you have given doesn't explain much about if this software works for structural learning with missing data.
@Marek Jozef Druzdzel, thank you for your reply , but I am looking for specifically structural learning with missing data. Your mentioned software talk about structural learning with no missing data, as you said.
@Irshad Samdani Mujawar
Replacing missing values or removing the records with missing values before structural learning solves the problem. I have done some research in the past that shows that the actual algorithm for treatment of missing data does not matter that much -- classifiers based on various algorithms show similar accuracy.
As others have mentioned, I would also recommend imputation. BnLearn in R can be used for Bayesian network analysis.
Irshad Samdani Mujawar the Pomegranate package in Python is very useful for BN. If you fill in the missing data points with NAN you can use learn the structure of the network as well as the CPTs based on the partial observations. I believe that the missing values are filled in with the most likely ones using EM and then the structure is learned based on the algorithm dictated by the user (exact, greedy, etc.)
Matlab. There is a BNT package for it:
https://github.com/bayesnet/bnt
- Badges
- Science method
- Similar topics
- Biological Science
- Ecology