Pairwise scores concept

[M Charles]

In order to compute pairwise scores for the edge feature Xi --> Xj, I believe we would use Proposition 3.2 of [Friedman & Koller, 2003] (although it looks like the i and j are flipped there). Here is a quick edited version:

formula.jpg (16.38 KiB) Viewed 1182 times

So, when computing the score of a particular node (Xj), all pairs that go from an earlier node to it should be pairwise scored.

As a side note, the paper mentions how calculating the probability of "a direct path from Xi to Xj" would have to look through exponentially many structures, so the original concept of pairwise scores would not work.