search

PoonLab / Kaphi

Kernel-embedded ABC-SMC for phylodynamic inference
GNU Affero General Public License v3.0
4 stars 2 forks source link

Possible for cherries to have label mismatch #40

Closed ArtPoon closed 7 years ago

ArtPoon commented 7 years ago

Currently when computing tree kernel at node pair n1 and n2, if left and right children of both nodes are tips then these subtrees rooted at n1 and n2 are a match. However, this situation may arise:

 A  B    B  A
 \  /    \  /
  n1      n2

This should still be counted as a match. Ladderizing the tree does not rotate cherries with respect to tip labels.

ArtPoon commented 7 years ago

Revisit this issue in light of recent work on Brad's project and stuff with Mariano

ArtPoon commented 7 years ago

Forgot that I refactored some of Brad's code that addresses this issue. Cherries are rotated to maximize kernel score. This is akin to ladderizing trees. Overall, the kernel score is comparing a set of tree shapes against another set, because some characteristics are ambiguous such as rotations around nodes. Ladderizing the trees deterministically collapses these sets, but some ambiguity remains at cherries. By rotating cherries to maximize the kernel score between two trees, we are resolving this ambiguity on a pair-by-pair basis. Is this still a valid kernel function?