discretization and parameter learning

[samoht]

Hello,

a few question to the discretization tool and mapping of nodes to column for parameter learning:

• I have decimal numbers in a column I want to discretize. But the automatically given bin names doesnt represent the bin borders correctly. E.g. s17_98_98 for bin borders from 98.3 to 98.7. So the bin borders are not reconstructable from the bin names. And for many columns it is too much effort to adapt the bin name for every bin manually. (e.g into s17_98_3_98_7)
• I want to map my discretized 60 bins to the node states. Is there another way than to create 60 states manually and name them before mapping?
• After mapping of nodes and saving the file sometimes some of the nodes are missing in the middle matching box of the window "Match network and data" after pressing ok and reopening the window. Why I cannot save the mapping?

Many thanks
Thomas

[shooltz[BayesFusion]]

I have decimal numbers in a column I want to discretize. But the automatically given bin names doesnt represent the bin borders correctly. E.g. s17_98_98 for bin borders from 98.3 to 98.7. So the bin borders are not reconstructable from the bin names. And for many columns it is too much effort to adapt the bin name for every bin manually. (e.g into s17_98_3_98_7)

What is the question?

I want to map my discretized 60 bins to the node states. Is there another way than to create 60 states manually and name them before mapping?

No, you need the manually entered names if you want 1:1 relationship between discretization intervals and node states. If you save your data as .gdat file, the discretization information is preserved and you'll be able to fine tune the intervals, for example.

After mapping of nodes and saving the file sometimes some of the nodes are missing in the middle matching box of the window "Match network and data" after pressing ok and reopening the window. Why I cannot save the mapping?

The feature is not implemented yet, but it's on the TODO list.

[samoht]

Hi,

many thanks for the fast reply.

I have decimal numbers in a column I want to discretize. But the automatically given bin names doesnt represent the bin borders correctly. E.g. s17_98_98 for bin borders from 98.3 to 98.7. So the bin borders are not reconstructable from the bin names. And for many columns it is too much effort to adapt the bin name for every bin manually. (e.g into s17_98_3_98_7)

What is the question?

My question here was if the automatically given names should not be better chosen so that the exact discretization interval borders are reconstructable like s17_98_3_98_7 for the intervall from 98.3 to 98.7 instead of s17_98_98. If I have 60 bins it is much work to adapt all names manually. Is there another way to do it?

Question by the way: why is there no undo-button in Genie? Always if i change something in a network i cannot undo it anymore.

Further question: If i have trained the parameters of a network with the data file and want to train the network with another datafile do i have to reset the trained parameters before or are the new result parameters not influenced by te start configuration?

best regards
Thomas

[shooltz[BayesFusion]]

My question here was if the automatically given names should not be better chosen so that the exact discretization interval borders are reconstructable like s17_98_3_98_7 for the intervall from 98.3 to 98.7 instead of s17_98_98. If I have 60 bins it is much work to adapt all names manually. Is there another way to do it?

The only way to automate it would be through SMILearn-based program (C++, Java, .Net)

Question by the way: why is there no undo-button in Genie? Always if i change something in a network i cannot undo it anymore.

Undo/redo is a nontrivial feature for BN editing program, quite labor-expensive to implement.

Further question: If i have trained the parameters of a network with the data file and want to train the network with another datafile do i have to reset the trained parameters before or are the new result parameters not influenced by te start configuration?

Check out the explanations in the GeNIe's EM dialog - screenshot attached, the relevant part marked red.