Jaccard score

[UnixJunkie]

Hello,

For binary fingerprints, you should probably use this version from scikit_learn:

jaccard_score(y_true, y_pred, average="micro")

Regards,
Francois.

[UnixJunkie]

Or, maybe even better: you write your own and can be completely sure of what is being computed.
J(A, B) = |AnB| / |AuB|

[beta-sheet]

Is there a particular reason for average="micro"? From the documentation it only seems to make a difference for data with more than 2 classes (whereas the fingerprints are binary).

jaccard_score works with the scikit-learn GaussianProcessRegressor, but I had to add a custom kernel for the GP implementation in PyTorch (which requires fingerprints be passed as torch.tensor).