About Parameter Learning

The front end.
Post Reply
niazmahmud
Posts: 3
Joined: Wed Jul 22, 2020 5:51 am

About Parameter Learning

Post by niazmahmud »

From a complete dataset, i used pc algorithm for structure learning (A DAG was formed). Then parameter learning was done. Then marginal probability was found. But when i further done parameter learning from the same data set, the marginal probability was changed. In example, first gender state 1 was 95% but after further parameter learning it was 5%. What is the reason? Step by step figure is attached below. Thanks in advances.
Attachments
Capture.PNG
Capture.PNG (84.22 KiB) Viewed 5384 times
Capture5.PNG
Capture5.PNG (255.98 KiB) Viewed 5384 times
Capture4.PNG
Capture4.PNG (152.48 KiB) Viewed 5384 times
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: About Parameter Learning

Post by shooltz[BayesFusion] »

Please post your data file and the network (after PC, before EM).
niazmahmud
Posts: 3
Joined: Wed Jul 22, 2020 5:51 am

Re: About Parameter Learning

Post by niazmahmud »

Here it is
Attachments
After_EM.xdsl
(6.85 KiB) Downloaded 341 times
Before_EM.xdsl
(6.85 KiB) Downloaded 314 times
modified_ROADSIDE_BBN.csv
(7.18 KiB) Downloaded 327 times
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: About Parameter Learning

Post by shooltz[BayesFusion] »

The problem probably is related to the automatic matching between data columns and network nodes. Your data file contains numeric indices. During structure learning these indices are converted into string outcome identifiers. The PC/EM does it internally and there's no issue. Later on, when you run EM again, your numeric data is not automatically matched to proper node outcomes - see the attached image and the yellow elements in the center list.

The easiest workaround would be to convert your data to labels (1 -> s1, 2 -> s2, etc). This approach will avoid the mismatch. Alternatively, you can drag'n'drop items in GeNIe's match network and data window, but this of course is a boring manual process.
Attachments
match-network-and-data.png
match-network-and-data.png (24.98 KiB) Viewed 5377 times
niazmahmud
Posts: 3
Joined: Wed Jul 22, 2020 5:51 am

Re: About Parameter Learning

Post by niazmahmud »

Ok. Thanks.
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: About Parameter Learning

Post by shooltz[BayesFusion] »

An update: the automatic matching between integer indices (1,2,...) and the corresponding state names (Xyz1, Xyz2, Xyz3...) has been added to GeNIe. We plan to release the next version of the program later this summer.
Post Reply