feat(skore): Support skrub dataops in reports

.github/workflows/sphinx.yml

@@ -230,6 +230,10 @@ jobs:
           path: skore/venv
           key: ${{ steps.cache-python-venv.outputs.cache-primary-key }}
 
+      - name: Install graphviz
+        run: |
+          sudo apt-get install -y graphviz
+
       - name: Install all packages from local branch
         run: |
           python -m pip install --force-reinstall --no-deps skore/

ci/requirements/skore/python-3.10/scikit-learn-1.5/test-requirements.txt

@@ -198,7 +198,7 @@ six==1.17.0
     # via python-dateutil
 skore-local-project==0.0.5
     # via skore (skore/pyproject.toml)
-skrub==0.7.2
+skrub==0.8.0
     # via skore (skore/pyproject.toml)
 stack-data==0.6.3
     # via ipython

ci/requirements/skore/python-3.10/scikit-learn-1.7/test-requirements.txt

@@ -198,7 +198,7 @@ six==1.17.0
     # via python-dateutil
 skore-local-project==0.0.5
     # via skore (skore/pyproject.toml)
-skrub==0.7.2
+skrub==0.8.0
     # via skore (skore/pyproject.toml)
 stack-data==0.6.3
     # via ipython

ci/requirements/skore/python-3.11/scikit-learn-1.5/test-requirements.txt

@@ -199,7 +199,7 @@ six==1.17.0
     # via python-dateutil
 skore-local-project==0.0.5
     # via skore (skore/pyproject.toml)
-skrub==0.7.2
+skrub==0.8.0
     # via skore (skore/pyproject.toml)
 stack-data==0.6.3
     # via ipython

ci/requirements/skore/python-3.11/scikit-learn-1.8/test-requirements.txt

@@ -199,7 +199,7 @@ six==1.17.0
     # via python-dateutil
 skore-local-project==0.0.5
     # via skore (skore/pyproject.toml)
-skrub==0.7.2
+skrub==0.8.0
     # via skore (skore/pyproject.toml)
 stack-data==0.6.3
     # via ipython

ci/requirements/skore/python-3.12/scikit-learn-1.5/test-requirements.txt

@@ -199,7 +199,7 @@ six==1.17.0
     # via python-dateutil
 skore-local-project==0.0.5
     # via skore (skore/pyproject.toml)
-skrub==0.7.2
+skrub==0.8.0
     # via skore (skore/pyproject.toml)
 stack-data==0.6.3
     # via ipython

ci/requirements/skore/python-3.12/scikit-learn-1.8/test-requirements.txt

@@ -199,7 +199,7 @@ six==1.17.0
     # via python-dateutil
 skore-local-project==0.0.5
     # via skore (skore/pyproject.toml)
-skrub==0.7.2
+skrub==0.8.0
     # via skore (skore/pyproject.toml)
 stack-data==0.6.3
     # via ipython

ci/requirements/skore/python-3.13/scikit-learn-1.5/test-requirements.txt

@@ -199,7 +199,7 @@ six==1.17.0
     # via python-dateutil
 skore-local-project==0.0.5
     # via skore (skore/pyproject.toml)
-skrub==0.7.2
+skrub==0.8.0
     # via skore (skore/pyproject.toml)
 stack-data==0.6.3
     # via ipython

ci/requirements/skore/python-3.13/scikit-learn-1.6/test-requirements.txt

@@ -199,7 +199,7 @@ six==1.17.0
     # via python-dateutil
 skore-local-project==0.0.5
     # via skore (skore/pyproject.toml)
-skrub==0.7.2
+skrub==0.8.0
     # via skore (skore/pyproject.toml)
 stack-data==0.6.3
     # via ipython

ci/requirements/skore/python-3.13/scikit-learn-1.7/test-requirements.txt

@@ -199,7 +199,7 @@ six==1.17.0
     # via python-dateutil
 skore-local-project==0.0.5
     # via skore (skore/pyproject.toml)
-skrub==0.7.2
+skrub==0.8.0
     # via skore (skore/pyproject.toml)
 stack-data==0.6.3
     # via ipython

ci/requirements/skore/python-3.13/scikit-learn-1.8/sphinx-requirements.txt

@@ -103,6 +103,8 @@ graphql-core==3.2.8
     #   graphql-relay
 graphql-relay==3.2.0
     # via graphene
+greenlet==3.3.2
+    # via sqlalchemy
 gunicorn==25.1.0
     # via mlflow
 h11==0.16.0
@@ -394,7 +396,7 @@ skore-local-project==0.0.5
     # via skore (skore/pyproject.toml)
 skore-mlflow-project==0.0.2
     # via skore (skore/pyproject.toml)
-skrub==0.7.2
+skrub==0.8.0
     # via skore (skore/pyproject.toml)
 smmap==5.0.3
     # via gitdb

ci/requirements/skore/python-3.13/scikit-learn-1.8/test-requirements.txt

@@ -199,7 +199,7 @@ six==1.17.0
     # via python-dateutil
 skore-local-project==0.0.5
     # via skore (skore/pyproject.toml)
-skrub==0.7.2
+skrub==0.8.0
     # via skore (skore/pyproject.toml)
 stack-data==0.6.3
     # via ipython

ci/requirements/skore/python-3.13/scikit-learn-1.8_and_pandas-2/test-requirements.txt

@@ -202,7 +202,7 @@ six==1.17.0
     # via python-dateutil
 skore-local-project==0.0.5
     # via skore (skore/pyproject.toml)
-skrub==0.7.2
+skrub==0.8.0
     # via skore (skore/pyproject.toml)
 stack-data==0.6.3
     # via ipython

examples/use_cases/plot_fraud_prediction.py

@@ -0,0 +1,94 @@
+"""
+Tracking all the data processing
+================================
+
+To track all operations and be able to apply the fitted estimator to unseen
+data, we need to include all the data wrangling in the estimator used for our
+skore report. In very simple cases this can be done with a scikit-learn
+Pipeline. When we have transformations not supported by the Pipeline (such as
+transformations that change the number of rows, or that involve multiple tables
+such as joins), skore allows us to use a skrub DataOp instead.
+
+In this example we consider a dataset that is simple, but still requires some
+data wrangling (encoding, aggregation and joining) which could not be performed
+in a regular scikit-learn estimator.
+
+To track those operations, we use a skrub DataOp, which can perform richer
+transformations than normal estimators, and also has built-in support from
+skore.
+
+The dataset contains a list of online transactions (each corresponds to a cart,
+or "basket"), each linked to one or more products for which we have a description.
+The task is to predict which involved credit fraud.
+"""
+
+# %%
+# We start by defining our data-processing pipeline. Note that it contains
+# operations, such as aggregating and joining the product information after
+# vectorizing the text it contains, that would not be possible in a normal
+# estimator.
+
+# %%
+import skore
+import skrub
+from sklearn.ensemble import HistGradientBoostingClassifier
+
+dataset = skrub.datasets.fetch_credit_fraud(split="all")
+
+products = skrub.var("products", dataset.products)
+baskets = skrub.var("baskets", dataset.baskets)
+
+basket_ids = baskets[["ID"]].skb.mark_as_X()
+fraud_flags = baskets["fraud_flag"].skb.mark_as_y()
+
+
+def filter_products(products, basket_ids):
+    return products[products["basket_ID"].isin(basket_ids["ID"])]
+
+
+vectorized_products = products.skb.apply_func(filter_products, basket_ids).skb.apply(
+    skrub.TableVectorizer(), exclude_cols="basket_ID"
+)
+
+
+def join_product_info(basket_ids, vectorized_products):
+    return basket_ids.merge(
+        vectorized_products.groupby("basket_ID").agg("mean").reset_index(),
+        left_on="ID",
+        right_on="basket_ID",
+    ).drop(columns=["ID", "basket_ID"])
+
+
+pred = basket_ids.skb.apply_func(join_product_info, vectorized_products).skb.apply(
+    HistGradientBoostingClassifier(), y=fraud_flags
+)
+
+# This would generate a report with previews of intermediate results & fitted
+# estimators:
+#
+# pred.skb.full_report()
+
+pred
+
+# %%
+# Above we see a preview on the whole dataset. Click the "show graph" toggle to
+# see a drawing of the pipeline we have built.
+#
+# Just like a normal estimator, a skrub DataOp can be used with skore reports.
+# We can either pass separately a SkrubLearner and training and testing data,
+# or pass our DataOp with the data it already contains and rely on the default
+# train/test split:
+
+# %%
+report = skore.EstimatorReport(pred, pos_label=1)
+report.metrics.roc_auc()
+
+# %%
+report.metrics.precision_recall().plot()
+
+# %%
+# Note that the preprocessing operations are captured in the skrub DataOp,
+# hence in our report -- so we can replay them later on unseen data.
+
+# %%
+report.estimator_.data_op

skore/pyproject.toml

@@ -35,7 +35,7 @@ dependencies = [
   "rich>=14.2",
   "scikit-learn",
   "skore[local]",
-  "skrub",
+  "skrub>=0.8",
   "seaborn",
   "jinja2",
 ]

skore/src/skore/_sklearn/_base.py

@@ -69,11 +69,11 @@ def _rich_repr(self, class_name: str) -> str:
         )
         return string_buffer.getvalue()
 
-    def _get_X_y(
+    def _get_data_and_y_true(
         self,
         *,
         data_source: Literal["test", "train"],
-    ) -> tuple[ArrayLike, ArrayLike]:
+    ) -> tuple[dict, ArrayLike]:
         """Get the requested dataset.
 
         Parameters
@@ -86,41 +86,31 @@ def _get_X_y(
 
         Returns
         -------
-        X : array-like of shape (n_samples, n_features)
+        data : dict of input data
             The requested dataset.
 
         y : array-like of shape (n_samples,)
-            The requested dataset.
+            The target labels.
         """
-        if data_source == "test":
-            if self._parent._X_test is None or self._parent._y_test is None:
-                missing_data = "X_test and y_test"
-                raise ValueError(
-                    f"No {data_source} data (i.e. {missing_data}) were provided "
-                    f"when creating the report. Please provide the {data_source} "
-                    "data when creating the report."
-                )
-            return self._parent._X_test, self._parent._y_test
-        elif data_source == "train":
-            if self._parent._X_train is None or self._parent._y_train is None:
-                missing_data = "X_train and y_train"
-                raise ValueError(
-                    f"No {data_source} data (i.e. {missing_data}) were provided "
-                    f"when creating the report. Please provide the {data_source} "
-                    "data when creating the report."
-                )
-            return self._parent._X_train, self._parent._y_train
-        else:
+        if data_source not in ["train", "test"]:
             raise ValueError(
                 f"Invalid data source: {data_source}. Possible values are: test, train."
             )
+        if getattr(self._parent, f"{data_source}_data") is None:
+            raise ValueError(
+                f"No {data_source} data were provided when creating the report."
+            )
+        if data_source == "test":
+            return self._parent.test_data, self._parent.y_test
+        assert data_source == "train"
+        return self._parent.train_data, self._parent.y_train
 
 
 def _get_cached_response_values(
     *,
     cache: Cache,
     estimator: BaseEstimator,
-    X: ArrayLike | None,
+    X: ArrayLike | dict | None,
     response_method: str | list[str] | tuple[str, ...],
     pos_label: PositiveLabel | None = None,
     data_source: Literal["test", "train"] = "test",

skore/src/skore/_sklearn/_comparison/report.py

@@ -455,14 +455,14 @@ def create_estimator_report(
 
         estimator_report = cast(EstimatorReport, self.reports_[report_key])
         X_concat = (
-            pd.concat([estimator_report._X_train, estimator_report._X_test])
-            if isinstance(estimator_report._X_train, pd.DataFrame)
-            else np.concatenate([estimator_report._X_train, estimator_report._X_test])
+            pd.concat([estimator_report.X_train, estimator_report.X_test])
+            if isinstance(estimator_report.X_train, pd.DataFrame)
+            else np.concatenate([estimator_report.X_train, estimator_report.X_test])
         )
         y_concat = (
-            pd.concat([estimator_report._y_train, estimator_report._y_test])
-            if isinstance(estimator_report._y_train, (pd.DataFrame, pd.Series))
-            else np.concatenate([estimator_report._y_train, estimator_report._y_test])
+            pd.concat([estimator_report.y_train, estimator_report.y_test])
+            if isinstance(estimator_report.y_train, (pd.DataFrame, pd.Series))
+            else np.concatenate([estimator_report.y_train, estimator_report.y_test])
         )
         report = EstimatorReport(
             estimator_report.estimator,