More efforts are needed to make the "Visualizing cross-validation behavior in skada" example easy to understand.
What is the interest of doing a DA cross-validation compared to using make_da_pipeline with scikit-learn cross-validation (like in "Using cross_val_score with skada" example)?
The LeaveOneDomainOut needs more explanations. It is not clear that domains are represented twice (for when they are considered source or target).
Are the test samples from source domains ever useful? From what I understand the adapter uses source and target train samples, then the base estimator is fitted with the source train samples and finally predict/score is performed on target test samples.
I made changes in the "Using cross_val_score with skada" example. I will now have a look at the two other examples of this section.