Fix/handle categorical features #30798 #30799
Conversation
The average_precision_score function was returning misleading results (1.0 or 0.0) when given a single sample. Added validation to require at least 2 samples and provide a clear error message. Fixes scikit-learn#30615
The average_precision_score function was returning misleading results (1.0 or 0.0) when given a single sample. Added validation to require at least 2 samples and provide a clear error message. Fixes scikit-learn#30615
The average_precision_score function was returning misleading results (1.0 or 0.0) when given a single sample. Added validation to require at least 2 samples and provide a clear error message. Fixes scikit-learn#30615
The average_precision_score function was returning misleading results (1.0 or 0.0) when given a single sample. Added validation to require at least 2 samples and provide a clear error message. Fixes scikit-learn#30615
…-sample-30615 Fix:Fix/average precision score single sample 30615
❌ Linting issuesThis PR is introducing linting issues. Here's a summary of the issues. Note that you can avoid having linting issues by enabling You can see the details of the linting issues under the
|
|
The original issue shows a stacktrace where the error is coming from xgboost, not scikit-learn, so I'm not sure what's to fix here. |
| if isinstance(X, pd.DataFrame): | ||
| X_new = X.iloc[:, candidate_mask] | ||
| else: | ||
| X_new = X[:, candidate_mask] |
There was a problem hiding this comment.
You should the function _safe_indexing instead.
However, here it is already to late because you already too late. We should avoid the call to validate_data or at least not validate X to not convert it to a NumPy array.
We also need to use n_features = _num_features(X) to compute the number of features.
There was a problem hiding this comment.
@glemaitre Should I modify the earlier validation step to prevent X from being converted to a NumPy array? If so, should I change how validate_data is called, or should we handle it differently?
There was a problem hiding this comment.
validate_data has a skip_check_array arg which would skip conversion to numpy.
Fix: Handle Categorical Features in
SequentialFeatureSelector([#30785](#30785))Overview:
This update resolves an issue in the
SequentialFeatureSelectorclass, where text and categorical features were not handled properly despite the estimator supporting them. The issue was tracked under [#30785](#30785).Problem:
The
SequentialFeatureSelectorwas not correctly handling pandas DataFrame inputs that included categorical or text-based features, leading to errors during feature selection. Although some estimators (e.g.,XGBRegressor) support categorical features, the feature selector was failing to process them as expected.Solution:
To address this issue, the following change was made to the
sklearn/feature_selection/_sequential.pyfile:Added Handling for Pandas DataFrames:
The code was updated to ensure compatibility with pandas DataFrames. When the input
Xis a DataFrame, it uses.iloc[]to correctly slice the columns based on thecandidate_mask.Impact:
SequentialFeatureSelectornow properly handles both numeric and categorical features in DataFrames.XGBRegressor, which support categorical data, now function correctly with the feature selector without raising errors.Testing: