It looks like a single test failed in one of the builds:

=================================== FAILURES ===================================
________________ test_leave_k_out_gets_correct_train_only_shape ________________

    def test_leave_k_out_gets_correct_train_only_shape():
        """Test that the correct number of users appear *only* in the train set."""
    
        mat = _get_matrix()
        train, test = leave_k_out_split(mat, K=1, train_only_size=0.8)
        train_only = ~np.isin(np.unique(train.tocoo().row), test.tocoo().row)
    
>       assert train_only.sum() == int(train.shape[0] * 0.8)
E       assert 81 == 80

Given that the tests are only failing in a single build and passing in all others, my guess is that a completely null row was present in the randomly generated matrix that caused it to be included in the training set in addition to all other randomly chosen users. I think the best solution to this is to add a check onto _get_matrix() that ensures that there aren't any completely zero rows.

Hmmm, even after making the fix to ensure always-populated rows, the test is still failing intermittently, and it's failing intermittently both for the random 100x100 sparse matrix as well as the newly-added fixed matrix. My guess is that this is some function of the combination of the candidate_mask with the train_only mask in which we are making an assumption that all users are eligible to be included in the test set when this is not the case, so more end up in the train set than we plan for?

Usually, the value for the number of users is only slightly off of the chosen value. @benfred would an almostEquals here with a delta of ~5 or so suffice?

@benfred Looks like this last failing test is addressed by #652, and some of the fixes in this PR are in effect duplicated by #653. Let me know if / how you'd like to proceed in fixing the functionality of leave_k_out_split, happy to help in any way I can.