Weisfeiler Lehman kernel fitting process

[tvayer]

Hello,

I'm using the grakel library (thank you for the very clear documentation and the gathering work !) in order to make classification. But i'm confused about some results I have regarding the Weisfeiler Lehman kernel.

For what I understand about the kernel is that there is no "learning" process : if we fit the kernel on the entire dataset or a subset we should have the same result about the pairwise similarities between the graphs.

However when I run the following code I'm not getting the same kernel at the end. First I fit_transform on all Mutag data getting a K_from_all kernel and then I select a subset (with respect to train and test indices) of this kernel.

I then compare with the same kernel but fitted on a small subset of the data (with respect to the train subset) and transformed on the test subset. I'm getting a K_from_small kernel which is different from the K_from_all kernel :

Did I miss details about the fitting procedure of the kernel ? (for the shortest path kernel I recover the same kernels)

Thank you very much

Titouan

[ysig]

Good morning (from this side of the world)!

This was a small error inside the Weisfeiler Lehman transform(), that had as a result that labels, that where not found at fit(), could be assigned a new label that was already assigned.

This and other changes can be found on branch 0.1a6, concerning this and other changes in grakel.

Thanks a lot for your feedback.

[tvayer]

All right thank you very much !