[doc] Brief introduction to base_score.

doc/parameter.rst

@@ -390,6 +390,8 @@ Specify the learning task and the corresponding learning objective. The objectiv
     disable the estimation, specify a real number argument.
   - For sufficient number of iterations, changing this value will not have too much effect.
 
+  See :doc:`/tutorials/intercept` for more info.
+
 * ``eval_metric`` [default according to objective]
 
   - Evaluation metrics for validation data, a default metric will be assigned according to objective (rmse for regression, and logloss for classification, `mean average precision` for ``rank:map``, etc.)

doc/tutorials/custom_metric_obj.rst

@@ -271,7 +271,8 @@ available in XGBoost:
 We use ``multi:softmax`` to illustrate the differences of transformed prediction.  With
 ``softprob`` the output prediction array has shape ``(n_samples, n_classes)`` while for
 ``softmax`` it's ``(n_samples, )``. A demo for multi-class objective function is also
-available at :ref:`sphx_glr_python_examples_custom_softmax.py`.
+available at :ref:`sphx_glr_python_examples_custom_softmax.py`. Also, see
+:doc:`/tutorials/intercept` for some more explanation.
 
 
 **********************

doc/tutorials/index.rst

@@ -30,4 +30,5 @@ See `Awesome XGBoost <https://github.com/dmlc/xgboost/tree/master/demo>`_ for mo
   input_format
   param_tuning
   custom_metric_obj
+  intercept
   privacy_preserving

doc/tutorials/intercept.rst

@@ -0,0 +1,104 @@
+#########
+Intercept
+#########
+
+.. versionadded:: 2.0.0
+
+Since 2.0.0, XGBoost supports estimating the model intercept (named ``base_score``)
+automatically based on targets upon training. The behavior can be controlled by setting
+``base_score`` to a constant value. The following snippet disables the automatic
+estimation:
+
+.. code-block:: python
+
+    import xgboost as xgb
+
+    reg = xgb.XGBRegressor()
+    reg.set_params(base_score=0.5)
+
+In addition, here 0.5 represents the value after applying the inverse link function. See
+the end of the document for a description.
+
+Other than the ``base_score``, users can also provide global bias via the data field
+``base_margin``, which is a vector or a matrix depending on the task. With multi-output
+and multi-class, the ``base_margin`` is a matrix with size ``(n_samples, n_targets)`` or
+``(n_samples, n_classes)``.
+
+.. code-block:: python
+
+    import xgboost as xgb
+    from sklearn.datasets import make_regression
+
+    X, y = make_regression()
+
+    reg = xgb.XGBRegressor()
+    reg.fit(X, y)
+    # Request for raw prediction
+    m = reg.predict(X, output_margin=True)
+
+    reg_1 = xgb.XGBRegressor()
+    # Feed the prediction into the next model
+    reg.fit(X, y, base_margin=m)
+    reg.predict(X, base_margin=m)
+
+
+It specifies the bias for each sample and can be used for stacking an XGBoost model on top
+of other models, see :ref:`sphx_glr_python_examples_boost_from_prediction.py` for a worked
+example. When ``base_margin`` is specified, it automatically overrides the ``base_score``
+parameter. If you are stacking XGBoost models, then the usage should be relatively
+straightforward, with the previous model providing raw prediction and a new model using
+the prediction as bias. For more customized inputs, users need to take extra care of the
+link function. Let :math:`F` be the model and :math:`g` be the link function, since
+``base_score`` is overridden when sample-specific ``base_margin`` is available, we will
+omit it here:
+
+.. math::
+
+   g(E[y_i]) = F(x_i)
+
+
+When base margin :math:`b` is provided, it's added to the raw model output :math:`F`:
+
+.. math::
+
+   g(E[y_i]) = F(x_i) + b_i
+
+and the output of the final model is:
+
+
+.. math::
+
+   g^{-1}(F(x_i) + b_i)
+
+Using the gamma deviance objective ``reg:gamma`` as an example, which has a log link
+function, hence:
+
+.. math::
+
+   \ln{(E[y_i])} = F(x_i) + b_i \\
+   E[y_i] = \exp{(F(x_i) + b_i)}
+
+As a result, if you are feeding outputs from models like GLM with a corresponding
+objective function, make sure the outputs are not yet transformed by the inverse link.
+
+In the case of ``base_score`` (intercept), it can be accessed through
+:py:meth:`~xgboost.Booster.save_config` after estimation. Unlike the ``base_margin``, the
+returned value represents a value after applying inverse link.  With logistic regression
+and the logit link function as an example, given the ``base_score`` as 0.5,
+:math:`g(intercept) = logit(0.5) = 0` is added to the raw model output:
+
+.. math::
+
+   E[y_i] = g^{-1}{(F(x_i) + g(intercept))}
+
+and 0.5 is the same as :math:`base_score = g^{-1}(0) = 0.5`. This is more intuitive if you
+remove the model and consider only the intercept, which is estimated before the model is
+fitted:
+
+.. math::
+
+   E[y] = g^{-1}{g(intercept))} \\
+   E[y] = intercept
+
+For some objectives like MAE, there are close solutions, while for others it's estimated
+with one step Newton method.

python-package/xgboost/core.py

@@ -785,7 +785,7 @@ def __init__(
                  so it doesn't make sense to assign weights to individual data points.
 
         base_margin :
-            Base margin used for boosting from existing model.
+            Global bias for each instance. See :doc:`/tutorials/intercept` for details.
         missing :
             Value in the input data which needs to be present as a missing value. If
             None, defaults to np.nan.

python-package/xgboost/sklearn.py

@@ -1012,7 +1012,7 @@ def fit(
         sample_weight :
             instance weights
         base_margin :
-            global bias for each instance.
+            Global bias for each instance. See :doc:`/tutorials/intercept` for details.
         eval_set :
             A list of (X, y) tuple pairs to use as validation sets, for which
             metrics will be computed.
@@ -1152,7 +1152,7 @@ def predict(
             When this is True, validate that the Booster's and data's feature_names are
             identical.  Otherwise, it is assumed that the feature_names are the same.
         base_margin :
-            Margin added to prediction.
+            Global bias for each instance. See :doc:`/tutorials/intercept` for details.
         iteration_range :
             Specifies which layer of trees are used in prediction.  For example, if a
             random forest is trained with 100 rounds.  Specifying ``iteration_range=(10,
@@ -1605,7 +1605,7 @@ def predict_proba(
             When this is True, validate that the Booster's and data's feature_names are
             identical.  Otherwise, it is assumed that the feature_names are the same.
         base_margin :
-            Margin added to prediction.
+            Global bias for each instance. See :doc:`/tutorials/intercept` for details.
         iteration_range :
             Specifies which layer of trees are used in prediction.  For example, if a
             random forest is trained with 100 rounds.  Specifying `iteration_range=(10,
@@ -1948,7 +1948,7 @@ def fit(
                 weights to individual data points.
 
         base_margin :
-            Global bias for each instance.
+            Global bias for each instance. See :doc:`/tutorials/intercept` for details.
         eval_set :
             A list of (X, y) tuple pairs to use as validation sets, for which
             metrics will be computed.