Bayesian Research -- Different networks with same initial parameter

The front end.
Post Reply
shutaozheng
Posts: 5
Joined: Wed Jul 04, 2018 8:10 am

Bayesian Research -- Different networks with same initial parameter

Post by shutaozheng »

Dear all,

I am a new Genie user and I have a doubt with Genie.
I tried to find the relationship(network) of some items by using Genie. I followed next steps:

1. I got my file in format csv.
2. Discretization of each row. There are 8 parameters (colums) and 590 rows.
3. Learn new network. I choose Bayesian Research, without change any initial parameters: discrete thresold: 20, iteration: 20, sample size: 50.
Initial Parameters- Bayesian Research
Initial Parameters- Bayesian Research
Captura2.PNG (21.58 KiB) Viewed 7987 times
4. And I got a network. (Attachment 2)
Network 1
Network 1
Captura3.PNG (43.46 KiB) Viewed 7987 times
The problem is, when I repeat step 1-3, without any change (same file, same disrectization, same parameters). However, I got another network in step 4. (Please see attachment 3)
Network 2
Network 2
Captura4.PNG (45.02 KiB) Viewed 7987 times
I have no idea why the network changes.

Any answer will be appreciated!!! Many thanks!!!
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: Bayesian Research -- Different networks with same initial parameter

Post by shooltz[BayesFusion] »

You're using random seed value of zero, which means that actual seed is taken from the system clock and a couple of other system values. To obtain repeatable results, use nonzero seed.
shutaozheng
Posts: 5
Joined: Wed Jul 04, 2018 8:10 am

Re: Bayesian Research -- Different networks with same initial parameter

Post by shutaozheng »

shooltz[BayesFusion] wrote:You're using random seed value of zero, which means that actual seed is taken from the system clock and a couple of other system values. To obtain repeatable results, use nonzero seed.
Thanks a lot!!! I put seed 1 and now I get the same network !

Hope you do not mind.... I have another question: as my object is to find the influence value of each parameter related with KEA(my last parameter). I re-constructed the network. I followed these steps:

1. Connect the arcs and build a new network as you can see in this picture.
network for influence
network for influence
New Network for influence 1.PNG (48.44 KiB) Viewed 7969 times
2. Learn Parameter : match the prameters and update the network.
Learn parameters
Learn parameters
Learn parameter.PNG (42.6 KiB) Viewed 7969 times
3. An automatic window is skipped: learn parameters with EM. I choose 'Keep Original' and Confidence '50' (the same number as sample size).
Learn parameters with EM
Learn parameters with EM
Learn parameter with EM.PNG (54.5 KiB) Viewed 7969 times
4. Finally I update the network and click on 'Strength of Influence' to get the influence value of each parameter related to KEA.

However, when I repeat again step 1-3 without change any input and click again on 'Strength of influence', the influence value of the parameters changes.

Do you know where the error is? (As Step 1 and 2 are only to match the parameters, I do not think there exist any incoherence, the problem may occur in Step 3).

Thanks in advance!!!
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: Bayesian Research -- Different networks with same initial parameter

Post by shooltz[BayesFusion] »

find the influence value of each parameter related with KEA(my last parameter)
Apply sensitivity analysis for this. See Using GeNIe | Bayesian Networks | Sensitivity analysis in Bayesian networks chapter in GeNIe manual.
shutaozheng
Posts: 5
Joined: Wed Jul 04, 2018 8:10 am

Re: Bayesian Research -- Different networks with same initial parameter

Post by shutaozheng »

shooltz[BayesFusion] wrote:
find the influence value of each parameter related with KEA(my last parameter)
Apply sensitivity analysis for this. See Using GeNIe | Bayesian Networks | Sensitivity analysis in Bayesian networks chapter in GeNIe manual.

Thanks very much for your prompt reply.

I also tried Sensitivity Analysis, but it changes every time when I rebuild the same network.

I attach two trials for your reference [Sensitivity Analysis & Strength of Influence]. All the previous steps are the same. but the final influence diagram result is different.

Trial 1: follow step 1 -3 I mencioned previously.
Influence distance of each parameter trial1
Influence distance of each parameter trial1
Trial 1.PNG (64.58 KiB) Viewed 7962 times
Trial 2: repite Trial 1 without any changes.
Influence distance of each parameter trial2
Influence distance of each parameter trial2
Trial 2.PNG (67.86 KiB) Viewed 7962 times
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: Bayesian Research -- Different networks with same initial parameter

Post by shooltz[BayesFusion] »

Don't mix strenght of influence and sensitivity analysis. Use one or the other. I recommend sensitivity analysis for your case.
shutaozheng
Posts: 5
Joined: Wed Jul 04, 2018 8:10 am

Re: Bayesian Research -- Different networks with same initial parameter

Post by shutaozheng »

shooltz[BayesFusion] wrote:Don't mix strenght of influence and sensitivity analysis. Use one or the other. I recommend sensitivity analysis for your case.
I tried Strength of influence firstly, the final influence diagram is different. Then I tried only Sensitivity Analysis, the result is not the same neither.

As the influence diagram varies every time I rebuilt the network (with same steps), I cannot sure the weight that each parameter contributes to the Target (In this case, KEA).
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: Bayesian Research -- Different networks with same initial parameter

Post by shooltz[BayesFusion] »

Post two networks learned with identical seed here, so we can investigate the problem.
shutaozheng
Posts: 5
Joined: Wed Jul 04, 2018 8:10 am

Re: Bayesian Research -- Different networks with same initial parameter

Post by shutaozheng »

shooltz[BayesFusion] wrote:Post two networks learned with identical seed here, so we can investigate the problem.
Dear shooltz,

I tried several times and I find the error! In step 3 I put sample size (confidence level) 50 instead of 100, it causes different results because Genie pick up randomly half of the samples (50%) to build the influence diagram.

when the sample size is 100, in all trials, my final Strength of Influence o Sensitiviy Analysis of each parameter get the same result.

Thanks a lot for your help!
Post Reply