Add support for uplifting in the inspector.

This allows exporting uplift models from YDF to TF

PiperOrigin-RevId: 579181978

Author: rstz

tensorflow_decision_forests/component/builder/builder.py

@@ -343,6 +343,10 @@ def _import_dataspec(self, src_dataspec: data_spec_pb2.DataSpecification):
         if dst_col_idx == self._header.ranking_group_col_idx:
           continue
 
+      if isinstance(self._objective, py_tree.objective.AbstractUpliftObjective):
+        if dst_col_idx == self._header.uplift_treatment_col_idx:
+          continue
+
       if not created:
         raise ValueError(
             "import_dataspec was called after some of the model was build. "
@@ -511,7 +515,11 @@ def _initialize_header_column_idx(self):
     label_column.name = self._objective.label
     self._dataspec_column_index[label_column.name] = self._header.label_col_idx
 
-    if isinstance(self._objective, py_tree.objective.ClassificationObjective):
+    if isinstance(
+        self._objective, py_tree.objective.ClassificationObjective
+    ) or isinstance(
+        self._objective, py_tree.objective.CategoricalUpliftObjective
+    ):
       label_column.type = ColumnType.CATEGORICAL
 
       # One value is reserved for the non-used OOV item.
@@ -537,6 +545,7 @@ def _initialize_header_column_idx(self):
         (
             py_tree.objective.RegressionObjective,
             py_tree.objective.RankingObjective,
+            py_tree.objective.NumericalUpliftObjective,
         ),
     ):
       label_column.type = ColumnType.NUMERICAL
@@ -556,6 +565,18 @@ def _initialize_header_column_idx(self):
           self._header.ranking_group_col_idx
       )
 
+    if isinstance(self._objective, py_tree.objective.AbstractUpliftObjective):
+      assert len(self._dataspec.columns) == 1
+
+      # Create the "treatment" column for Uplifting.
+      self._header.uplift_treatment_col_idx = 1
+      treatment_column = self._dataspec.columns.add()
+      treatment_column.type = ColumnType.CATEGORICAL
+      treatment_column.name = self._objective.treatment
+      self._dataspec_column_index[treatment_column.name] = (
+          self._header.uplift_treatment_col_idx
+      )
+
 
 @six.add_metaclass(abc.ABCMeta)
 class AbstractDecisionForestBuilder(AbstractBuilder):
@@ -850,8 +871,11 @@ def check_leaf(self, node: py_tree.node.LeafNode):
             "A regression objective requires leaf nodes with regressive values."
         )
 
-    elif isinstance(self.objective, py_tree.objective.RankingObjective):
-      raise ValueError("Ranking objective not supported by this model")
+    elif isinstance(self.objective, py_tree.objective.AbstractUpliftObjective):
+      if not isinstance(node.value, py_tree.value.UpliftValue):
+        raise ValueError(
+            "An uplift objective requires leaf nodes with uplift values."
+        )
 
     else:
       raise NotImplementedError()
@@ -920,6 +944,11 @@ def __init__(
       loss = gradient_boosted_trees_pb2.Loss.LAMBDA_MART_NDCG5
       bias = [bias]
 
+    elif isinstance(objective, py_tree.objective.AbstractUpliftObjective):
+      raise ValueError(
+          "Uplift objective not supported by Gradient Boosted Tree models."
+      )
+
     else:
       raise NotImplementedError()
 
@@ -972,7 +1001,12 @@ def specialized_header_filename(self) -> str:
     return self._file_prefix + inspector_lib.BASE_FILENAME_GBT_HEADER
 
   def check_leaf(self, node: py_tree.node.LeafNode):
-    if not isinstance(node.value, py_tree.value.RegressionValue):
+    if isinstance(self.objective, py_tree.objective.AbstractUpliftObjective):
+      raise ValueError(
+          "Uplift objective not supported by Gradient Boosted Tree models."
+      )
+
+    elif not isinstance(node.value, py_tree.value.RegressionValue):
       raise ValueError(
           "A GBT model should only have leaf with regressive "
           f"value. Got {node.value} instead."

tensorflow_decision_forests/component/inspector/inspector.py

@@ -252,6 +252,22 @@ def objective(self) -> py_tree.objective.AbstractObjective:
       return py_tree.objective.RankingObjective(
           label=label.name, group=group_column.name)
 
+    elif self.task == Task.CATEGORICAL_UPLIFT:
+      uplift_treatment = self._dataspec.columns[
+          self._header.uplift_treatment_col_idx
+      ]
+      return py_tree.objective.CategoricalUpliftObjective(
+          label=label.name, treatment=uplift_treatment.name
+      )
+
+    elif self.task == Task.NUMERICAL_UPLIFT:
+      uplift_treatment = self._dataspec.columns[
+          self._header.uplift_treatment_col_idx
+      ]
+      return py_tree.objective.NumericalUpliftObjective(
+          label=label.name, treatment=uplift_treatment.name
+      )
+
     else:
       raise NotImplementedError()
 

tensorflow_decision_forests/component/py_tree/objective.py

@@ -155,3 +155,46 @@ def __eq__(self, other):
     if not isinstance(other, RankingObjective):
       return False
     return self.label == other.label and self._group == other._group
+
+
+class AbstractUpliftObjective(AbstractObjective):
+  """Objective for Uplift."""
+
+  def __init__(self, label: str, treatment: str):
+    super(AbstractUpliftObjective, self).__init__(label)
+    self._treatment = treatment
+
+  @property
+  def treatment(self) -> str:
+    return self._treatment
+
+  def __eq__(self, other):
+    if not isinstance(other, AbstractUpliftObjective):
+      return False
+    return (
+        self.label == other.label
+        and self._treatment == other._treatment
+        and self.task == other.task
+    )
+
+
+class CategoricalUpliftObjective(AbstractUpliftObjective):
+  """Objective for Categorical Uplift."""
+
+  @property
+  def task(self) -> Task:
+    return Task.CATEGORICAL_UPLIFT
+
+  def __repr__(self):
+    return f"CategoricalUplift(label={self.label}, treatment={self._treatment})"
+
+
+class NumericalUpliftObjective(AbstractUpliftObjective):
+  """Objective for Numerical Uplift."""
+
+  @property
+  def task(self) -> Task:
+    return Task.NUMERICAL_UPLIFT
+
+  def __repr__(self):
+    return f"NumericalUplift(label={self.label}, treatment={self._treatment})"

tensorflow_decision_forests/component/py_tree/objective_test.py

@@ -66,6 +66,27 @@ def test_ranking(self):
     objective = objective_lib.RankingObjective(label="label", group="group")
     logging.info("objective: %s", objective)
 
+  def test_numerical_uplift(self):
+    objective = objective_lib.NumericalUpliftObjective(
+        label="label", treatment="treatment"
+    )
+    logging.info("objective: %s", objective)
+
+  def test_categorical_uplift(self):
+    objective = objective_lib.CategoricalUpliftObjective(
+        label="label", treatment="treatment"
+    )
+    logging.info("objective: %s", objective)
+
+  def test_uplift_objects_are_not_equal(self):
+    numerical_objective = objective_lib.NumericalUpliftObjective(
+        label="label", treatment="treatment"
+    )
+    categorical_objective = objective_lib.CategoricalUpliftObjective(
+        label="label", treatment="treatment"
+    )
+    self.assertNotEqual(numerical_objective, categorical_objective)
+
 
 if __name__ == "__main__":
   tf.test.main()

tensorflow_decision_forests/keras/core_inference.py

@@ -1239,7 +1239,6 @@ def yggdrasil_model_to_keras_model(
         "file containing among other things, a data_spec.pb file."
     )
 
-  temp_directory = None
   if src_container == "zip":
     # Unzip the model in a temporary directory
     temp_directory = tempfile.TemporaryDirectory()
@@ -1255,6 +1254,13 @@ def yggdrasil_model_to_keras_model(
       ranking_group=objective.group
       if objective.task == inspector_lib.Task.RANKING
       else None,
+      uplift_treatment=objective.treatment
+      if objective.task
+      in (
+          inspector_lib.Task.CATEGORICAL_UPLIFT,
+          inspector_lib.Task.NUMERICAL_UPLIFT,
+      )
+      else None,
       verbose=verbose,
       advanced_arguments=AdvancedArguments(
           disable_categorical_integer_offset_correction=disable_categorical_integer_offset_correction,

tensorflow_decision_forests/keras/keras_test.py

@@ -2293,7 +2293,9 @@ def test_golden_model_gbt(self):
       ("adult_binary_class_gbdt", 0.012131),
       ("prefixed_adult_binary_class_gbdt", 0.012131),
   )
-  def test_ydf_to_keras_model(self, ydf_model_directory, expected_prediction):
+  def test_ydf_to_keras_model_adult(
+      self, ydf_model_directory, expected_prediction
+  ):
     ygg_model_path = os.path.join(
         ydf_test_data_path(), "model", ydf_model_directory
     )
@@ -2328,6 +2330,28 @@ def custom_model_input_signature(
     )
     self.assertNear(prediction[0, 0], expected_prediction, 0.00001)
 
+  def test_ydf_to_keras_model_uplift(self):
+    ygg_model_path = os.path.join(
+        ydf_test_data_path(), "model", "sim_pte_categorical_uplift_rf"
+    )
+    tfdf_model_path = os.path.join(tmp_path(), "sim_pte_categorical_uplift_rf")
+
+    dataset_directory = os.path.join(ydf_test_data_path(), "dataset")
+    test_path = os.path.join(dataset_directory, "sim_pte_test.csv")
+    test_df = pd.read_csv(test_path)
+
+    outcome_key = "y"
+    treatment_group = "treat"
+    # Remove the treatment group from the test dataset.
+    test_df = test_df.drop(treatment_group, axis=1)
+
+    core.yggdrasil_model_to_keras_model(ygg_model_path, tfdf_model_path)
+    loaded_model = models.load_model(tfdf_model_path)
+    prediction = loaded_model.predict(
+        keras.pd_dataframe_to_tf_dataset(test_df, label=outcome_key)
+    )
+    self.assertNear(prediction[0, 0], -0.7580058, 0.00001)
+
   @parameterized.parameters(
       "directory",
       "zip",