Validating Network with GeNIe

The front end.
Post Reply
oscarpc
Posts: 33
Joined: Mon Jun 04, 2012 1:17 am

Validating Network with GeNIe

Post by oscarpc »

Hello all,

I have got some doubts that might be very simple, but they are giving me some headache.

I have built a DBN and its parameters have been automatically learned with a dataset. This DBN has only two time slices and the nodes that participate in the "time series" are called DM (DM for slice 0 and DM_1 for slice 1) and SoP (SoP for slice 0 and SoP _1 for slice 1) . SoP is a hidden variable.

I have ONLY observations (node Sodium_0) in slice 0 and I would like to predict DM_0 in slice 1. That is, for training purposes, I have used the data available for each slice. For prediction purposes, I would like to consider observations only in slice 0.

My questions are:

1. If I wish to predict what happens in a node, DM_1 for example, do I have to define DM_1 as a "target node"?
2. How can I indicate to the network that the observations (Sodium) are only available in slice 0?

Finally, I am trying to use the option of GeNIe Data->Validation, but I am afraid I don't know how to use it. It gives me an error message for each record in the test dataset that says: "Can't instantiate data record #(record number)".

Some questions with regard to this menu:
3. When the program asks for the time series variables, what do I have to put here? All time series nodes in order or only the one I would like to predict?
4. In addition, the screen "Match Network and Data" shows DM_1 and SoP_1 in the right hand side column. Does this mean that they haven't been associated? That happens to me every time, that is, it seems that the temporal nodes only associate the first slice.
Can we learn parameters and validate the network for Dynamic Bayesian Networks in GeNIe?

I have included a simplified version of my network, the training and test data. I have also included a Dummy Variable column at the end of my data. If I don't do that, the column DM_1 is not accessible at all.

Thank you very much for all your help.

Kind regards,
Oscar PC.
Attachments
data_1dayhosp_v01_testing.txt
(6.43 KiB) Downloaded 358 times
data_1dayhosp_v01_training.txt
(19.99 KiB) Downloaded 369 times
DBN.xdsl.zip
(99.1 KiB) Downloaded 326 times
oscarpc
Posts: 33
Joined: Mon Jun 04, 2012 1:17 am

Re: Validating Network with GeNIe

Post by oscarpc »

Hello,
Does anyone know any of the answers to my previous questions :-)?
Thank you.
Oscar
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: Validating Network with GeNIe

Post by shooltz[BayesFusion] »

Can we learn parameters and validate the network for Dynamic Bayesian Networks in GeNIe?
I believe I have answered some of your questions in this thread, but I don't see my post now. In short, you can learn DBN parameters in GeNIe. You can't validate DBN, unless you unroll it first to obtain non-dynamic network. Are you comfortable with the UI for network validation when applied to non-dynamic networks?
oscarpc
Posts: 33
Joined: Mon Jun 04, 2012 1:17 am

Re: Validating Network with GeNIe

Post by oscarpc »

Thank you very much.
It was really useful.

One last two questions, I have read some posts about "target nodes", but still, I don't have very clear when it is useful to define one node as "target" and the other as "targets".

In addition, is there any algorithm (i.e. Viterbi) in Genie that gives me the most probable sequence of states in a node, including hidden nodes (no observations)?

As always, thank you so much for your kind and always useful help.
Kind regards,
Oscar PC.
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: Validating Network with GeNIe

Post by shooltz[BayesFusion] »

I don't have very clear when it is useful to define one node as "target" and the other as "targets".
I'm not sure I understand your questions ("target" vs. "targets" part).

By selecting one or more explicit targets you may reduce the complexity of the inference workload. With target nodes present the algorithm calculates the beliefs of target nodes only. If no explicit targets are selected, all beliefs are calculated.
In addition, is there any algorithm (i.e. Viterbi) in Genie that gives me the most probable sequence of states in a node, including hidden nodes (no observations)?
Inference in DBNs will produce beliefs for all timeslices - you just need to iterate over the array and select the most probable state in each slice.
Martijn
Posts: 76
Joined: Sun May 29, 2011 12:23 am

Re: Validating Network with GeNIe

Post by Martijn »

We have the Annealed MAP algorithm implemented which will find the most probable sequence of states, but it doesn't work directly on DBNs, i.e. with temperal plates. Currently it crashes.
It will work when you unroll your network.

You can find it under Network -> Annealed Map
Right clicking on the node allows you to either set evidence or add it to the MAP (to have it's most probable state determined).

Best,

Martijn
oscarpc
Posts: 33
Joined: Mon Jun 04, 2012 1:17 am

Re: Validating Network with GeNIe

Post by oscarpc »

Thanks Martijn,

Sorry, I meant "target" versus "non target" nodes. I forgot to put the "non" :P
Thank you so much for your help, very useful. I will try the Annealed Map algorithm.

One more question. I don't quite understand what the difference is between having the DBN unrolled or not.
Does it make any difference to the parameter learning process?

Thank you very much.
Kind regards.

Oscar PC
Martijn
Posts: 76
Joined: Sun May 29, 2011 12:23 am

Re: Validating Network with GeNIe

Post by Martijn »

Hi Oscar,
One more question. I don't quite understand what the difference is between having the DBN unrolled or not.
Does it make any difference to the parameter learning process?
No, it's just a matter of representation.
A DBN is in fact just a normal Bayesian network, with some added symantics (we model the passage of time in a certain way).
The "normal" DBN view using the temporal plate is just a very compact view of the DBN, when we unroll the network we show the actual Bayesian network (I believe that under the hood we do inference and learning with the unrolled network).
Due to implementation specifics not all the features of GeNIe operate properly on the compact DBN view, but will work on the unrolled network.

Martijn
oscarpc
Posts: 33
Joined: Mon Jun 04, 2012 1:17 am

Re: Validating Network with GeNIe

Post by oscarpc »

Thank you Martijn, I really appreciate your help.
Kind regards,

Oscar PC.
Post Reply