Return error when training a node classification model on negative integer labels.

dgf/src/learning/ten_lines/node_prediction_test.py

@@ -996,6 +996,64 @@ def test_label_classes_fails_for_integer_labels(self):
     self.assertIn("does not have a string dictionary", str(ctx.exception))
 
 
+class NodePredictionNegativeLabelTest(absltest.TestCase):
+  """Integer classification labels must be non-negative.
+
+  A negative label (e.g. -1) generally encodes a missing label.
+  """
+
+  @classmethod
+  def setUpClass(cls):
+    super().setUpClass()
+    cls.graph, cls.schema = gen_test_graph.gen_toy_classification_dataset()
+    # The first 10 nodes have a missing label.
+    cls.graph.node_sets["N1"].features["label"][:10] = -1
+
+  def test_in_memory_negative_label_raises(self):
+    with self.assertRaisesRegex(
+        ValueError, "should be non-negative. However, its minimum value is -1"
+    ):
+      node_prediction_lib.train_node_model(
+          graph=self.graph,
+          schema=self.schema,
+          target_nodeset="N1",
+          target_column="label",
+          **RAPID_TRAINING_KWARGS,
+      )
+
+  def test_pre_sampled_negative_label_raises(self):
+    sampler = in_memory_sampler_lib.create_sampler(
+        graph=self.graph,
+        plan=sampling_config_lib.simple_sampling_config_to_sampling_plan(
+            sampling_config_lib.SimpleSamplingConfig(
+                seed_nodeset="N1", num_hops=1, hop_width=3, reverse=True
+            ),
+            self.schema,
+        ),
+        schema=self.schema,
+        batch_size=10,
+    )
+    path = os.path.join(self.create_tempdir().full_path, "samples@2.tfrecord")
+    tf_graph_sample_lib.write_tfgnn_graphs(
+        (sample for sample in sampler.sample(np.arange(10))),
+        path,
+        schema=self.schema,
+        container_type="TF_RECORD",
+    )
+
+    with self.assertRaisesRegex(
+        ValueError, "should be non-negative. However, its minimum value is -1"
+    ):
+      node_prediction_lib.train_node_model(
+          graph=path,
+          valid_graph=path,
+          schema=self.schema,
+          target_nodeset="N1",
+          target_column="label",
+          **RAPID_TRAINING_KWARGS,
+      )
+
+
 class NodePredictionRegressionToy(parameterized.TestCase):
 
   @parameterized.parameters(("float32", 1), ("int64", 1), ("float32", 3))

dgf/src/learning/ten_lines/node_prediction_train.py

@@ -466,6 +466,23 @@ def train_node_model(
             f" '{task.target_nodeset}' must have `num_categorical_values`"
             " defined for classification tasks."
         )
+      if (
+          schema.node_sets[task.target_nodeset]
+          .features[task.target_column]
+          .format.is_integer()
+      ):
+        label_stats = dataset_preparator.feature_stats.node_sets[
+            task.target_nodeset
+        ].features[task.target_column]
+        if label_stats.minimum < 0:
+          raise ValueError(
+              f"Target column '{task.target_column}' in nodeset"
+              f" '{task.target_nodeset}' is an integer categorical feature, so"
+              " its values should be non-negative. However, its minimum value"
+              f" is {label_stats.minimum:g}. Negative values generally indicate"
+              " missing labels. Make sure that all the nodes have a valid"
+              " label."
+          )
     elif task.task_type == node_prediction_model.TaskType.NODE_REGRESSION:
       if label_spec.semantic != schema_lib.FeatureSemantic.NUMERICAL:
         raise ValueError(