Pushing the docs to dev/ for branch: main, commit 89cf3d24e8ff892e14d611fdabc852928f46e27b

Author: sklearn-ci

dev/_downloads/02192f99342d6d1323161978b6c80bfc/plot_ols_ridge.zip

@@ -141,7 +141,7 @@
       "cell_type": "markdown",
       "metadata": {},
       "source": [
-        "Due to the asymmetry of the distribution of the noise, we observe that the\ntrue mean and estimated conditional median are different. We also observe\nthat each quantile model has different parameters to better fit the desired\nquantile. Note that ideally, all quantiles would be parallel in this case,\nwhich would become more visible with more data points or less extreme\nquantiles, e.g. 10% and 90%.\n\n## Comparing `QuantileRegressor` and `LinearRegression`\n\nIn this section, we will linger on the difference regarding the error that\n:class:`~sklearn.linear_model.QuantileRegressor` and\n:class:`~sklearn.linear_model.LinearRegression` are minimizing.\n\nIndeed, :class:`~sklearn.linear_model.LinearRegression` is a least squares\napproach minimizing the mean squared error (MSE) between the training and\npredicted targets. In contrast,\n:class:`~sklearn.linear_model.QuantileRegressor` with `quantile=0.5`\nminimizes the mean absolute error (MAE) instead.\n\nLet's first compute the training errors of such models in terms of mean\nsquared error and mean absolute error. We will use the asymmetric Pareto\ndistributed target to make it more interesting as mean and median are not\nequal.\n\n"
+        "Due to the asymmetry of the distribution of the noise, we observe that the\ntrue mean and estimated conditional median are different. We also observe\nthat each quantile model has different parameters to better fit the desired\nquantile. Note that ideally, all quantiles would be parallel in this case,\nwhich would become more visible with more data points or less extreme\nquantiles, e.g. 10% and 90%.\n\n## Comparing `QuantileRegressor` and `LinearRegression`\n\nIn this section, we will linger on the difference regarding the loss functions that\n:class:`~sklearn.linear_model.QuantileRegressor` and\n:class:`~sklearn.linear_model.LinearRegression` are minimizing.\n\nIndeed, :class:`~sklearn.linear_model.LinearRegression` is a least squares\napproach minimizing the mean squared error (MSE) between the training and\npredicted targets. In contrast,\n:class:`~sklearn.linear_model.QuantileRegressor` with `quantile=0.5`\nminimizes the mean absolute error (MAE) instead.\n\nWhy does it matter? The loss functions specify what exactly the model is aiming\nto predict, see\n`user guide on the choice of scoring function<which_scoring_function>`.\nIn short, a model minimizing MSE predicts the mean (expectation) and a model\nminimizing MAE predicts the median.\n\nLet's compute the training errors of such models in terms of mean\nsquared error and mean absolute error. We will use the asymmetric Pareto\ndistributed target to make it more interesting as mean and median are not\nequal.\n\n"
       ]
     },
     {
@@ -152,7 +152,7 @@
       },
       "outputs": [],
       "source": [
-        "from sklearn.linear_model import LinearRegression\nfrom sklearn.metrics import mean_absolute_error, mean_squared_error\n\nlinear_regression = LinearRegression()\nquantile_regression = QuantileRegressor(quantile=0.5, alpha=0)\n\ny_pred_lr = linear_regression.fit(X, y_pareto).predict(X)\ny_pred_qr = quantile_regression.fit(X, y_pareto).predict(X)\n\nprint(\n    f\"\"\"Training error (in-sample performance)\n    {linear_regression.__class__.__name__}:\n    MAE = {mean_absolute_error(y_pareto, y_pred_lr):.3f}\n    MSE = {mean_squared_error(y_pareto, y_pred_lr):.3f}\n    {quantile_regression.__class__.__name__}:\n    MAE = {mean_absolute_error(y_pareto, y_pred_qr):.3f}\n    MSE = {mean_squared_error(y_pareto, y_pred_qr):.3f}\n    \"\"\"\n)"
+        "from sklearn.linear_model import LinearRegression\nfrom sklearn.metrics import mean_absolute_error, mean_squared_error\n\nlinear_regression = LinearRegression()\nquantile_regression = QuantileRegressor(quantile=0.5, alpha=0)\n\ny_pred_lr = linear_regression.fit(X, y_pareto).predict(X)\ny_pred_qr = quantile_regression.fit(X, y_pareto).predict(X)\n\nprint(\n    \"Training error (in-sample performance)\\n\"\n    f\"{'model':<20}   MAE   MSE\\n\"\n    f\"{linear_regression.__class__.__name__:<20} \"\n    f\"{mean_absolute_error(y_pareto, y_pred_lr):5.3f} \"\n    f\"{mean_squared_error(y_pareto, y_pred_lr):5.3f}\\n\"\n    f\"{quantile_regression.__class__.__name__:<20} \"\n    f\"{mean_absolute_error(y_pareto, y_pred_qr):5.3f} \"\n    f\"{mean_squared_error(y_pareto, y_pred_qr):5.3f}\"\n)"
       ]
     },
     {
@@ -170,7 +170,7 @@
       },
       "outputs": [],
       "source": [
-        "from sklearn.model_selection import cross_validate\n\ncv_results_lr = cross_validate(\n    linear_regression,\n    X,\n    y_pareto,\n    cv=3,\n    scoring=[\"neg_mean_absolute_error\", \"neg_mean_squared_error\"],\n)\ncv_results_qr = cross_validate(\n    quantile_regression,\n    X,\n    y_pareto,\n    cv=3,\n    scoring=[\"neg_mean_absolute_error\", \"neg_mean_squared_error\"],\n)\nprint(\n    f\"\"\"Test error (cross-validated performance)\n    {linear_regression.__class__.__name__}:\n    MAE = {-cv_results_lr[\"test_neg_mean_absolute_error\"].mean():.3f}\n    MSE = {-cv_results_lr[\"test_neg_mean_squared_error\"].mean():.3f}\n    {quantile_regression.__class__.__name__}:\n    MAE = {-cv_results_qr[\"test_neg_mean_absolute_error\"].mean():.3f}\n    MSE = {-cv_results_qr[\"test_neg_mean_squared_error\"].mean():.3f}\n    \"\"\"\n)"
+        "from sklearn.model_selection import cross_validate\n\ncv_results_lr = cross_validate(\n    linear_regression,\n    X,\n    y_pareto,\n    cv=3,\n    scoring=[\"neg_mean_absolute_error\", \"neg_mean_squared_error\"],\n)\ncv_results_qr = cross_validate(\n    quantile_regression,\n    X,\n    y_pareto,\n    cv=3,\n    scoring=[\"neg_mean_absolute_error\", \"neg_mean_squared_error\"],\n)\nprint(\n    \"Test error (cross-validated performance)\\n\"\n    f\"{'model':<20}   MAE   MSE\\n\"\n    f\"{linear_regression.__class__.__name__:<20} \"\n    f\"{-cv_results_lr['test_neg_mean_absolute_error'].mean():5.3f} \"\n    f\"{-cv_results_lr['test_neg_mean_squared_error'].mean():5.3f}\\n\"\n    f\"{quantile_regression.__class__.__name__:<20} \"\n    f\"{-cv_results_qr['test_neg_mean_absolute_error'].mean():5.3f} \"\n    f\"{-cv_results_qr['test_neg_mean_squared_error'].mean():5.3f}\"\n)"
       ]
     },
     {

dev/_downloads/02fe21a1f5d14bd1ebbe34aa905d9bf6/plot_multilabel.zip

@@ -231,7 +231,7 @@
 # Comparing `QuantileRegressor` and `LinearRegression`
 # ----------------------------------------------------
 #
-# In this section, we will linger on the difference regarding the error that
+# In this section, we will linger on the difference regarding the loss functions that
 # :class:`~sklearn.linear_model.QuantileRegressor` and
 # :class:`~sklearn.linear_model.LinearRegression` are minimizing.
 #
@@ -241,7 +241,13 @@
 # :class:`~sklearn.linear_model.QuantileRegressor` with `quantile=0.5`
 # minimizes the mean absolute error (MAE) instead.
 #
-# Let's first compute the training errors of such models in terms of mean
+# Why does it matter? The loss functions specify what exactly the model is aiming
+# to predict, see
+# :ref:`user guide on the choice of scoring function<which_scoring_function>`.
+# In short, a model minimizing MSE predicts the mean (expectation) and a model
+# minimizing MAE predicts the median.
+#
+# Let's compute the training errors of such models in terms of mean
 # squared error and mean absolute error. We will use the asymmetric Pareto
 # distributed target to make it more interesting as mean and median are not
 # equal.
@@ -255,14 +261,14 @@
 y_pred_qr = quantile_regression.fit(X, y_pareto).predict(X)
 
 print(
-    f"""Training error (in-sample performance)
-    {linear_regression.__class__.__name__}:
-    MAE = {mean_absolute_error(y_pareto, y_pred_lr):.3f}
-    MSE = {mean_squared_error(y_pareto, y_pred_lr):.3f}
-    {quantile_regression.__class__.__name__}:
-    MAE = {mean_absolute_error(y_pareto, y_pred_qr):.3f}
-    MSE = {mean_squared_error(y_pareto, y_pred_qr):.3f}
-    """
+    "Training error (in-sample performance)\n"
+    f"{'model':<20}   MAE   MSE\n"
+    f"{linear_regression.__class__.__name__:<20} "
+    f"{mean_absolute_error(y_pareto, y_pred_lr):5.3f} "
+    f"{mean_squared_error(y_pareto, y_pred_lr):5.3f}\n"
+    f"{quantile_regression.__class__.__name__:<20} "
+    f"{mean_absolute_error(y_pareto, y_pred_qr):5.3f} "
+    f"{mean_squared_error(y_pareto, y_pred_qr):5.3f}"
 )
 
 # %%
@@ -294,14 +300,14 @@
     scoring=["neg_mean_absolute_error", "neg_mean_squared_error"],
 )
 print(
-    f"""Test error (cross-validated performance)
-    {linear_regression.__class__.__name__}:
-    MAE = {-cv_results_lr["test_neg_mean_absolute_error"].mean():.3f}
-    MSE = {-cv_results_lr["test_neg_mean_squared_error"].mean():.3f}
-    {quantile_regression.__class__.__name__}:
-    MAE = {-cv_results_qr["test_neg_mean_absolute_error"].mean():.3f}
-    MSE = {-cv_results_qr["test_neg_mean_squared_error"].mean():.3f}
-    """
+    "Test error (cross-validated performance)\n"
+    f"{'model':<20}   MAE   MSE\n"
+    f"{linear_regression.__class__.__name__:<20} "
+    f"{-cv_results_lr['test_neg_mean_absolute_error'].mean():5.3f} "
+    f"{-cv_results_lr['test_neg_mean_squared_error'].mean():5.3f}\n"
+    f"{quantile_regression.__class__.__name__:<20} "
+    f"{-cv_results_qr['test_neg_mean_absolute_error'].mean():5.3f} "
+    f"{-cv_results_qr['test_neg_mean_squared_error'].mean():5.3f}"
 )
 
 # %%