diff --git a/hlink/linking/core/pipeline.py b/hlink/linking/core/pipeline.py
index e06edc8..8331a31 100644
--- a/hlink/linking/core/pipeline.py
+++ b/hlink/linking/core/pipeline.py
@@ -12,6 +12,7 @@
     Interaction,
 )
 import hlink.linking.transformers.float_cast_transformer
+from hlink.linking.transformers.rename_vector_attributes import RenameVectorAttributes
 import logging
 
 logger = logging.getLogger(__name__)
@@ -143,7 +144,25 @@ def generate_pipeline_stages(conf, ind_vars, tf, tconf):
                         inputCols=input_cols,
                         outputCol=pipeline_feature["output_column"],
                     )
+
+                    # Spark's Interaction creates its output vector attribute names
+                    # by concatenating the input column names with colons :. This
+                    # works fine for most of the down-pipeline transformers, but
+                    # LightGBM cannot run with attribute names that contain colons.
+                    # So this custom hlink transformer replaces colons in the vector
+                    # attribute names with underscores.
+                    #
+                    # Without this step, the colons propagate into the attribute
+                    # names for the features vector created by the VectorAssembler
+                    # and cause an error when training a LightGBM model.
+                    remove_colons_from_interaction_vector = RenameVectorAttributes(
+                        inputCol=interaction.getOutputCol(),
+                        strsToReplace=[":"],
+                        replaceWith="_",
+                    )
+
                     pipeline_stages.append(interaction)
+                    pipeline_stages.append(remove_colons_from_interaction_vector)
 
     if len(categorical_pipeline_features) > 0:
         encoded_output_cols = [