Directed Models: Limitations
n
Acyclicity constraint limits
expressive power:
n
Two objects linked to by a
student probably not both
professors
n
n
Allow arbitrary patterns over
sets of objects & links
n
Acyclicity forces modeling
of all potential links:
n
Network size
O(N
2
)
n
Inference is
quadratic
n
Generative training:
n
Train to fit all of data, not
to maximize accuracy
n
Influence flows over existing
links, exploiting link graph
sparsity
n
Network size
O(N)
n
Allow discriminative training:
n
Max P (labels | observations)
Solution: Undirected Models
[Lafferty, McCallum, Pereira]