Feature importance in data drift

How to show feature importance in Data Drift evaluations.

You are looking at the old Evidently documentation: this API is available with versions 0.6.7 or lower. Check the newer docs version here.

You can add feature importances to the dataset-level data drift Tests and Metrics:

• DataDriftTable

• TestShareOfDriftedColumns

Code example

Notebook example on showing feature importance:

evidently/examples/how_to_questions/how_to_add_feature_importances_to_drift.ipynb at ad71e132d59ac3a84fce6cf27bd50b12b10d9137 · evidentlyai/evidentlyGitHub

Compute feature importances

By default, the feature importance column is not shown. To display them, you must set the feature_importance parameter as True.

report = Report(metrics = [
    DataDriftTable(feature_importance=True)
])

If you do not specify anything else, Evidently will train a random forest model using the provided dataset and derive the feature importances.

Notes:

• This is only possible if your dataset contains the target column.

• If you have both current and reference datasets, two different models will be trained. You will have two columns with feature importance: one for reference and one for current data.

• If your dataset also contains the prediction column, you should clearly label it using Column Mapping to avoid it being treated as a feature.

Pass your own importances

You can also pass the list of feature importances derived during the model training process. This is a recommended option.

In this case, pass it as a list using the additional_data parameter when running the Report.

report = Report(metrics = [
    DataDriftTable(feature_importance=True)
])
report.run(reference_data=reference,
           current_data=current.loc['2011-01-29 00:00:00':'2011-02-07 23:00:00'],
           column_mapping=column_mapping,
           additional_data = {'current_feature_importance':
              dict(map(lambda i,j : (i,j), numerical_features + categorical_features, regressor.feature_importances_))
            }
           )

You can pass the current_feature_importance – a single column will appear in this case. You can also optionally pass reference_feature_importance.