|
|
 | | From: | Yaroslav Bulatov | | Subject: | Increasing complexity of a Graphical Model | | Date: | Tue, 18 Jan 2005 01:35:41 GMT |
|
|
 | Hi, I'm looking for a way to increase the complexity of graphical model
one parameter at a time. The problem is that when I add full n'th order
interaction, I add a CPT with j^n-1 parameters, so that's a big jump in
complexity.
More formally, the problem is the following. Start with binary feature
vector v= and learn a probability distribution f:v->[0..1]
from data. To make this tractable, I start with fully disconnected
Graphical Model. Then I add some (hyper)edge, retrain, look at cross
validation error, and repeat. When cross-validation accuracy stops
increasing, terminate.
After I allowed all possible 4th order interactions, I may add a
hyperedge for 5th order interactions. But specifying that hyperedge as
a table will require me to specify 32 elements. That's a big jump in
complexity. What would be a cogent way to increase complexity one
parameter at a time? What assumptions does it make?
PS: A typical domain for this problem is Co-NLL named entity
recognition dataset.
Yaroslav Bulatov
bulatov@cs.oregonstate.edu
Dearborn 102
Oregon State University
Corvallis, OR
[ comp.ai is moderated. To submit, just post and be patient, or if ]
[ that fails mail your article to , and ]
[ ask your news administrator to fix the problems with your system. ]
|
|
|
