quantifying sensitivity to data,diff. in the resultin BNs

The engine.
Post Reply
frito
Posts: 32
Joined: Wed Dec 12, 2007 8:51 pm

quantifying sensitivity to data,diff. in the resultin BNs

Post by frito »

Is there a methodology or some kind of a metric to compare two networks?
(except for goodness of fit to data\likelihood score...this may not
reflect the changes I have in mind, see next)**


That is a measure of how much the network structure, parent child and
vice versa relations etc. has changed.


e.g. used to be: ....A->B.., new net: ..A<-B.. or A->...C..->B how to
quantify this?

Lets say I used slightly different data, or somehow normalized, or
missing values were replaced by mean value... and wanna see how
much is the learning of the network sensitive to such changes and
wanna quantify that.

**Back to the goodness of fit to data: two structurally completely
different networks can have the same error score, but it doesnt tell
me anything about how the structure's changed...


thx
mark
Posts: 179
Joined: Tue Nov 27, 2007 4:02 pm

Post by mark »

Can't you just count the number of edges that have been added, removed, and reversed? The (possibly weighted) total could be interpreted as a measure of difference.
Post Reply