[MRG] BUG apply sample weights to RANSAC loss functions #15952
Conversation
sklearn/linear_model/_ransac.py
Outdated
| @@ -373,7 +373,10 @@ def fit(self, X, y, sample_weight=None): | |||
|
|
|||
| # residuals of all data for current random sample model | |||
| y_pred = base_estimator.predict(X) | |||
| residuals_subset = loss_function(y, y_pred) | |||
| if sample_weight is None: | |||
There was a problem hiding this comment.
Could you use the sklearn.utils._check_sample_weight. It will give something like:
sample_weight = _check_sample_weigth(sample_weight, X)
residuals_subset = loss_function(y, y_pred) * sample_weight|
Thinking a bit more about I have the intuition that this is different from other models where we try to reduce the loss but taking into account all samples meaning we will try to reduce the error of the points with high weights. Here the high weights samples will only be discarded and recognized as outliers. @jnothman Does what I say make sense (or is it too late here :))? |
|
Hmmmm... Might have to think more about this when I have headspace to do
so.
I think that low weight on an outlier should mean it's more outlying... If
we ensure that the invariances we have previously discussed are upheld,
then this should be the case. But it might be hard to uphold them since
aspects of this algorithm are discrete?
|
|
I just updated the function to use the syntax suggested by @glemaitre. Though I don't have as thorough knowledge of sklearn, I think @glemaitre's point makes sense. Perhaps it would be more consistent to do something like the below? |
|
A non-regression test might check for specific values, which we've tested conform with our expectations in relation to master??? In any case, please add a |
|
Sure thing @jnothman, I can get to adding a test this week. I'll ping you once it's added 👍 |
|
Superseded by #23371 |
Reference Issues/PRs
Fixes #15836
What does this implement/fix? Explain your changes.
Ensures that
sample_weightis applied to the output of theloss_functionwhen determining inliers/outliers at each iteration.Any other comments?
I'm not too sure of the best way to add a non-regression test for this behavior. I originally tried fitting two RANSAC estimators with/without weights and then compared their output coefficients (which were indeed different) but there was no way to know that the difference was from
sample_weightbeing used during the underlying estimators'.fit()or if it was because ofsample_weightbeing applied to the output of theloss_function.Would it be useful to run an example with/without weights, checking the resulting
.inlier_mask_(or something similar), and then writing a test that asserts the.inlier_mask_is the same?