Training with a completely hidden node

The front end.
Post Reply
MartinA
Posts: 7
Joined: Thu Nov 29, 2018 2:51 pm

Training with a completely hidden node

Post by MartinA » Fri Nov 30, 2018 1:20 pm

Hi all,

I would like to have more insights on training a BN with a node that is completely hidden. Can I do that?

A small illustration to make things clearer:

I am building a network for the purpose of monitoring some variations in the performances of an electronic circuit. These variations cannot be extracted directly, but I know that they are correlated to sensors that I can measure. My strategy is then to build an intermediate hidden node V representing these abstract variations and link it to the values given by the sensors (and the circuit performances).

In essence, V is completely unknown during training (= composed of only missing instances). My current strategy is to include V in the network and give e.g. 5 possible values that it can take, but I'm not sure I can do that. In the dataset, V is a column composed of only missing instances. But when I assign the nodes and dataset columns in the "Match Network and Data" window, it then assigns State0 to the value 0 (that is never given in the dataset, all instances for V are missing) and the four other states at given at (unassigned).

After training, what I see is that among the 5 values that V can take, the probability of V=state0 is almost 1, and of course the probabilities of all other states are nearly zero.

Am I doing something wrong ? Or is this not handled/ not possible to be done?

Thanks in advance!

shooltz[BayesFusion]
Site Admin
Posts: 1181
Joined: Mon Nov 26, 2007 5:51 pm

Re: Training with a completely hidden node

Post by shooltz[BayesFusion] » Fri Nov 30, 2018 2:29 pm

EM in GeNIe does not require that every node has an associated data column. GeNIe will attempt to match as many node/column pairs as possible, based on the identifiers.

If V in the dataset has only missing values, either remove the column V, or drag the V node to the left list to ensure it is not associated with the V column in the dataset (dragging V column to the right will also work).

MartinA
Posts: 7
Joined: Thu Nov 29, 2018 2:51 pm

Re: Training with a completely hidden node

Post by MartinA » Fri Nov 30, 2018 3:48 pm

Oh actually that was easy :)

Thanks a lot it works well now !

Post Reply