BUG - Fixing a few edge cases in the generation of the parallel coordinate plot #1800
Conversation
rcap107
left a comment
There was a problem hiding this comment.
I am not sure the solution I am proposing is correct.
I was not able to reproduce the same problem that triggered the CI in the first place. np.uniform should handle the case in which min_value == max_value, but it fails if either is np.nan or np.inf.
However, the presence of np.nan should have been handled in _prepare_numeric_column, which replaces nans with a placeholder value.
I don't know if this bug is related to the fact that it was the optuna example.
|
It seems like the problem occurs if the column is integer and all values are np.nan |
|
In the end, the problem was in the function |
jeromedockes
left a comment
There was a problem hiding this comment.
great! thanks for the fix @rcap107
Bug Fix Pull Request
This PR addresses a bug that may occur when using
plot_parallel_coordif the results of the crossvalidation are not correct.Description
The PR addresses the problem by adding a small eps and removing nan values from what is used in
_add_jitter.How Has This Been Tested?
A new test has been added to
test_parallel_coord.pyto check the values that are passed to_add_jitterSteps to Reproduce (original bug)
The bug first appeared in the CI testing the optuna example:
https://app.circleci.com/pipelines/github/skrub-data/skrub/7618/workflows/eef20e21-776e-42d2-b24a-3031c0c05614/jobs/14823