feat: Display the model coefficients for linear models

[sylvaincom]

Introduces a new accessor on EstimatorReport, feature_importance, which currently has one method, coefficients, only available for estimators with a coef_ attribute (such as linear models).

Demo:

from sklearn.datasets import make_regression
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LinearRegression
X, y = make_regression(random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42)
estimator = LinearRegression().fit(X_train, y_train)
from skore import EstimatorReport
report = EstimatorReport(estimator, X_test=X_test, y_test=y_test)
report.help()
╭─────────────────── Tools to diagnose estimator LinearRegression ───────────────────╮
│ EstimatorReport                                                                      │
│ ├── .metrics                                                                         │
│ │   ├── .accuracy(...)         (↗︎)     - Compute the accuracy score.                 │
│ │   ├── .brier_score(...)      (↘︎)     - Compute the Brier score.                    │
│ │   ├── .log_loss(...)         (↘︎)     - Compute the log loss.                       │
│ │   ├── .precision(...)        (↗︎)     - Compute the precision score.                │
│ │   ├── .precision_recall(...)         - Plot the precision-recall curve.            │
│ │   ├── .recall(...)           (↗︎)     - Compute the recall score.                   │
│ │   ├── .roc(...)                      - Plot the ROC curve.                         │
│ │   ├── .roc_auc(...)          (↗︎)     - Compute the ROC AUC score.                  │
│ │   ├── .custom_metric(...)            - Compute a custom metric.                    │
│ │   └── .report_metrics(...)           - Report a set of metrics for our estimator.  │
│ ├── .feature_importance                                                              │
│ │   └── .coefficients(...)             - Report the coefficients of a regression     │
│ │       estimator.                                                                   │
│ ├── .cache_predictions(...)            - Cache estimator's predictions.              │
│ ├── .clear_cache(...)                  - Clear the cache.                            │
│ └── Attributes                                                                       │
│     ├── .X_test                                                                      │
│     ├── .y_test                                                                      │
│     ├── .estimator_                                                                  │
│     └── .estimator_name_                                                             │
│                                                                                      │
│                                                                                      │
│ Legend:                                                                              │
│ (↗︎) higher is better (↘︎) lower is better                                             │
╰──────────────────────────────────────────────────────────────────────────────────────╯

report.feature_importance.coefficients()

Note what happens if the estimator is not a regressor (e.g. a LogisticRegression):

report.feature_importance.help()
╭─ Available feature importance methods ─╮
│ report.feature_importance              │
╰────────────────────────────────────────╯

---

Close #1320

Co-authored-by: @auguste-probabl [email protected]

[github-actions]

Coverage Report for backend

File	Stmts	Miss	Cover	Missing
venv/lib/python3.12/site-packages/skore
__init__.py	15	0	100%
__main__.py	8	8	0%	3–19
_config.py	28	0	100%
exceptions.py	4	4	0%	4–23
venv/lib/python3.12/site-packages/skore/persistence
__init__.py	0	0	100%
venv/lib/python3.12/site-packages/skore/persistence/item
__init__.py	56	3	93%	96–99
altair_chart_item.py	19	1	91%	14
item.py	22	1	95%	86
matplotlib_figure_item.py	36	1	95%	19
media_item.py	22	0	100%
numpy_array_item.py	27	1	94%	16
pandas_dataframe_item.py	29	1	94%	14
pandas_series_item.py	29	1	94%	14
pickle_item.py	22	0	100%
pillow_image_item.py	25	1	93%	15
plotly_figure_item.py	20	1	92%	14
polars_dataframe_item.py	27	1	94%	14
polars_series_item.py	22	1	92%	14
primitive_item.py	23	2	91%	13–15
sklearn_base_estimator_item.py	29	1	94%	15
skrub_table_report_item.py	10	1	86%	11
venv/lib/python3.12/site-packages/skore/persistence/repository
__init__.py	2	0	100%
item_repository.py	59	5	91%	15–16, 202–203, 226
venv/lib/python3.12/site-packages/skore/persistence/storage
__init__.py	4	0	100%
abstract_storage.py	22	0	100%
disk_cache_storage.py	33	1	95%	44
in_memory_storage.py	20	0	100%
venv/lib/python3.12/site-packages/skore/project
__init__.py	3	0	100%
_open.py	5	0	100%
project.py	81	1	99%	284
venv/lib/python3.12/site-packages/skore/sklearn
__init__.py	6	0	100%
_base.py	162	13	92%	43, 115, 118, 171–180, 192–>197, 212, 215–216
find_ml_task.py	61	0	99%	136–>144
types.py	13	2	85%	33, 61
venv/lib/python3.12/site-packages/skore/sklearn/_comparison
__init__.py	5	0	100%
metrics_accessor.py	164	2	97%	165, 166–>168, 1218
precision_recall_curve_display.py	73	1	97%	196–>199, 304
prediction_error_display.py	67	10	78%	97, 154–>exit, 209, 214–218, 227, 231, 236–238
report.py	64	1	96%	16, 251–>254
roc_curve_display.py	69	1	96%	204–>213, 213–>216, 308
venv/lib/python3.12/site-packages/skore/sklearn/_cross_validation
__init__.py	5	0	100%
metrics_accessor.py	170	0	99%	142–>144, 144–>146
report.py	105	1	98%	22
venv/lib/python3.12/site-packages/skore/sklearn/_estimator
__init__.py	7	0	100%
feature_importance_accessor.py	39	0	100%
metrics_accessor.py	325	11	95%	166–175, 203–>212, 211, 241, 252–>254, 282, 309–313, 328, 351, 363, 364–>366
report.py	127	1	97%	22, 229–>235, 237–>239
venv/lib/python3.12/site-packages/skore/sklearn/_plot
__init__.py	4	0	100%
precision_recall_curve.py	129	1	98%	240–>257, 329
prediction_error.py	102	1	98%	173, 189–>192
roc_curve.py	143	0	100%
style.py	14	0	100%
utils.py	99	5	94%	31, 55–57, 61
venv/lib/python3.12/site-packages/skore/sklearn/train_test_split
__init__.py	0	0	100%
train_test_split.py	36	2	94%	16–17
venv/lib/python3.12/site-packages/skore/sklearn/train_test_split/warning
__init__.py	8	0	100%
high_class_imbalance_too_few_examples_warning.py	17	1	90%	79
high_class_imbalance_warning.py	18	0	100%
random_state_unset_warning.py	12	1	88%	15
shuffle_true_warning.py	10	1	83%	46
stratify_is_set_warning.py	12	1	88%	15
time_based_column_warning.py	23	2	86%	17, 73
train_test_split_warning.py	5	1	80%	21
venv/lib/python3.12/site-packages/skore/utils
__init__.py	6	0	100%
_accessor.py	17	0	100%
_environment.py	27	27	0%	1–51
_index.py	5	0	100%
_logger.py	22	22	0%	3–38
_parallel.py	38	3	88%	23–33, 124
_patch.py	13	5	53%	21–37
_progress_bar.py	34	0	100%
_show_versions.py	33	0	100%
TOTAL	2991	152	94%

Tests	Skipped	Failures	Errors	Time
636	3 💤	0 ❌	0 🔥	48.235s ⏱️

[github-actions]

Documentation preview @ 9f2734d

[glemaitre]

We should anticipate how hard is to do the plotting with only pandas. It is to decide whether or not we should have a display.

[sylvaincom]

We have a toy example for example-driven dev at examples/model_evaluation/plot_feature_importance.py

[MarieSacksick]

I checked /examples/model_evaluation/plot_feature_importance.py and from this, it seems nice to me!
I didn't check the code nor several different models.

[sylvaincom]

@thomass-dev do you know why the documentation preview does not include the new example? 🙏 Is it because the example was added after the documentation preview was generated?

[MarieSacksick]

Only one comment, otherwise it's good for me!

[glemaitre]

I just started by reviewing the example.

Now I'm thinking that .model_weights() should return a display. The plan would be that this display has .plot() to make the right plotting for the user, as .frame property that return the current dataframe.

In addition, I like the current direction for the pandas style and I think that we can make it the default HTML repr of the display. One issue with the style is that it does not return a real dataframe that you can manipulate. That's why, I think that we can use it in the repr since it is not an object that is used by the user afterwards and we still have the .frame in case someone wants to access the real dataframe.

[glemaitre]

We need to extend the test and check what happens when passing a classifier. It tells me that somehow, we might want to validate the shape of coef_ and intercept_ to something that we know how to handle and raise an error otherwise.

[auguste-probabl]

We need to extend the test and check what happens when passing a classifier. It tells me that somehow, we might want to validate the shape of coef_ and intercept_ to something that we know how to handle and raise an error otherwise.

We have a test that passes a LogisticRegression. Our strategy for dealing with the non-standardized shapes of coef_ and intercept_ has been to use np.atleast_2d, which seems to work pretty well.

[glemaitre]

We have a test that passes a LogisticRegression. Our strategy for dealing with the non-standardized shapes of coef_ and intercept_ has been to use np.atleast_2d, which seems to work pretty well.

Your test does not check for multiclass. As mentioned #1339 (comment), I think that the transpose in this case will do the trick.