[GSProcessing&GSPartition] Support Hard Negative

docs/source/advanced/link-prediction.rst

@@ -236,6 +236,8 @@ impact is negligible.
 
 With DGL 1.0.4, ``fast_localuniform`` dataloader can speedup 2.4X over ``localuniform`` dataloader on training a 2 layer RGCN on MAG dataset on four g5.48x instances.
 
+.. _hard_negative_sampling:
+
 Hard Negative sampling
 -----------------------
 GraphStorm provides support for users to define hard negative edges for a positive edge during Link Prediction training.
@@ -272,10 +274,10 @@ In general, GraphStorm covers following cases:
 Preparing graph data for hard negative sampling
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-The gconstruct pipeline of GraphStorm provides support to load hard negative data from raw input.
+Both single machine and distributed graph construction pipeline of GraphStorm provide support to load hard negative data from raw input.
 Hard destination negatives can be defined through ``edge_dst_hard_negative`` transformation.
 The ``feature_col`` field of ``edge_dst_hard_negative`` must stores the raw node ids of hard destination nodes.
-The follwing example shows how to define a hard negative feature for edges with the relation ``(node1, relation1, node1)``:
+The following example shows how to define a hard negative feature for edges with the relation ``(node1, relation1, node1)``:
 
   .. code-block:: json
 

docs/source/cli/graph-construction/distributed/gsprocessing/input-configuration.rst

@@ -491,6 +491,13 @@ arguments.
         You can use a length greater than the dataset's longest sentence; or for a safe value choose 128. Make sure to check
         the model's max supported length when setting this value.
 
+-  ``edge_dst_hard_negative``
+
+   -  Encodes a hard negative edge feature for link prediction. For detail information for hard negative support, please refer to :ref:`hard_negative_sampling`.
+   -  ``kwargs``:
+      - ``separator`` (String, optional): The separator is used to
+        split multiple values in an input string for data in CSV files e.g. ``p0;s1``. If it is not provided, then the whole value
+        will be treated as a single string.
 
 ..  _gsprocessing-multitask-ref:
 

graphstorm-processing/graphstorm_processing/config/config_conversion/gconstruct_converter.py

@@ -188,6 +188,12 @@ def _convert_feature(feats: list[Mapping[str, Any]]) -> list[dict]:
                         "hf_model": gconstruct_transform_dict["bert_model"],
                         "max_seq_length": gconstruct_transform_dict["max_seq_length"],
                     }
+                elif gconstruct_transform_dict["name"] == "edge_dst_hard_negative":
+                    gsp_transformation_dict["name"] = "edge_dst_hard_negative"
+                    if "separator" in gconstruct_transform_dict:
+                        gsp_transformation_dict["kwargs"] = {
+                            "separator": gconstruct_transform_dict["separator"],
+                        }
                 else:
                     raise ValueError(
                         "Unsupported GConstruct transformation name: "

graphstorm-processing/graphstorm_processing/config/config_parser.py

@@ -29,6 +29,7 @@
 )
 from .categorical_configs import MultiCategoricalFeatureConfig
 from .hf_configs import HFConfig
+from .hard_negative_configs import HardEdgeNegativeConfig
 from .data_config_base import DataStorageConfig
 
 
@@ -71,6 +72,8 @@ def parse_feat_config(feature_dict: Dict) -> FeatureConfig:
         return MultiCategoricalFeatureConfig(feature_dict)
     elif transformation_name == "huggingface":
         return HFConfig(feature_dict)
+    elif transformation_name == "edge_dst_hard_negative":
+        return HardEdgeNegativeConfig(feature_dict)
     else:
         raise RuntimeError(f"Unknown transformation name: '{transformation_name}'")
 

graphstorm-processing/graphstorm_processing/config/hard_negative_configs.py

@@ -0,0 +1,35 @@
+"""
+Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License").
+You may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
+from typing import Mapping
+
+from .feature_config_base import FeatureConfig
+
+
+class HardEdgeNegativeConfig(FeatureConfig):
+    """Feature configuration for hard negative feature. Now only support link prediction.
+
+    Supported kwargs
+    ----------------
+    separator: str, optional
+        The separator for string input value. Only required when input value type is string.
+    """
+
+    def __init__(self, config: Mapping):
+        super().__init__(config)
+        self.separator = self._transformation_kwargs.get("separator", None)
+
+        self._sanity_check()

graphstorm-processing/graphstorm_processing/constants.py

@@ -58,6 +58,14 @@
 HUGGINGFACE_TOKENIZE = "tokenize_hf"
 HUGGINGFACE_EMB = "embedding_hf"
 
+################# Hard Negative transformations  ################
+ORDER_INDEX = "hard_negative_order_id"
+EXPLODE_HARD_NEGATIVE_VALUE = "hard_negative_exploded_single_value"
+
+################# Node Mapping  ################
+NODE_MAPPING_STR = "orig"
+NODE_MAPPING_INT = "new"
+
 
 ################# Supported execution envs  ##############
 class ExecutionEnv(Enum):

graphstorm-processing/graphstorm_processing/data_transformations/dist_feature_transformer.py

@@ -28,6 +28,7 @@
     DistCategoryTransformation,
     DistMultiCategoryTransformation,
     DistHFTransformation,
+    DistHardEdgeNegativeTransformation,
 )
 
 
@@ -69,6 +70,10 @@ def __init__(
             self.transformation = DistMultiCategoryTransformation(**default_kwargs, **args_dict)
         elif feat_type == "huggingface":
             self.transformation = DistHFTransformation(**default_kwargs, **args_dict)
+        elif feat_type == "edge_dst_hard_negative":
+            self.transformation = DistHardEdgeNegativeTransformation(
+                **default_kwargs, **args_dict, spark=spark
+            )
         else:
             raise NotImplementedError(
                 f"Feature {feat_name} has type: {feat_type} that is not supported"

graphstorm-processing/graphstorm_processing/data_transformations/dist_transformations/__init__.py

@@ -15,3 +15,4 @@
 )
 from .dist_bucket_numerical_transformation import DistBucketNumericalTransformation
 from .dist_hf_transformation import DistHFTransformation
+from .dist_hard_negative_transformation import DistHardEdgeNegativeTransformation

graphstorm-processing/graphstorm_processing/data_transformations/dist_transformations/dist_hard_negative_transformation.py

@@ -0,0 +1,136 @@
+"""
+Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License").
+You may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
+from typing import Sequence
+from pyspark.sql.functions import split, col
+from pyspark.sql.types import ArrayType, IntegerType, StringType
+from pyspark.sql import DataFrame, functions as F, SparkSession
+
+from graphstorm_processing.constants import (
+    NODE_MAPPING_STR,
+    NODE_MAPPING_INT,
+    ORDER_INDEX,
+    EXPLODE_HARD_NEGATIVE_VALUE,
+)
+
+from .base_dist_transformation import DistributedTransformation
+
+
+def apply_transform(
+    cols: Sequence[str],
+    separator: str,
+    spark: SparkSession,
+    input_df: DataFrame,
+    hard_node_mapping_dict: dict,
+) -> DataFrame:
+    """Applies hard negative transformation to each row.
+
+    Parameters
+    ----------
+    cols : Sequence[str]
+        List of column names to apply normalization to.
+    separator: str, optional
+        The separator for string input value. Only required when input value type is string.
+    spark: SparkSession
+        The spark session.
+    input_df : DataFrame
+        The input DataFrame to apply transformation to.
+    hard_node_mapping_dict: dict
+        The mapping dictionary contain mapping file directory and edge type.
+    """
+    column_type = input_df.schema[cols[0]].dataType
+    if isinstance(column_type, StringType):
+        transformed_df = input_df.withColumn(cols[0], split(col(cols[0]), separator))
+    else:
+        transformed_df = input_df
+    # Edge type should be (src_ntype:relation_type:dst_ntype)
+    # Only support hard negative for destination nodes. Get the node type of destination nodes.
+    # TODO: support hard negative for source nodes.
+    _, _, dst_type = hard_node_mapping_dict["edge_type"].split(":")
+    mapping_prefix = hard_node_mapping_dict["mapping_path"]
+    format_name = hard_node_mapping_dict["format_name"]
+    hard_negative_node_mapping = spark.read.parquet(
+        f"{mapping_prefix}{dst_type}/{format_name}/*.parquet"
+    )
+    node_mapping_length = hard_negative_node_mapping.count()
+
+    # TODO: Use panda series to possibly improve the efficiency
+    # Explode the original list and join node id mapping dataframe
+    transformed_df = transformed_df.withColumn(ORDER_INDEX, F.monotonically_increasing_id())
+    transformed_df = transformed_df.withColumn(
+        EXPLODE_HARD_NEGATIVE_VALUE, F.explode(F.col(cols[0]))
+    )
+    transformed_df = transformed_df.join(
+        hard_negative_node_mapping,
+        transformed_df[EXPLODE_HARD_NEGATIVE_VALUE] == hard_negative_node_mapping[NODE_MAPPING_STR],
+        "inner",
+    ).select(NODE_MAPPING_INT, ORDER_INDEX)
+    transformed_df = transformed_df.groupBy(ORDER_INDEX).agg(
+        F.collect_list(NODE_MAPPING_INT).alias(cols[0])
+    )
+
+    # Same length for feature to convert to tensor
+    def pad_mapped_values(hard_neg_list):
+        while len(hard_neg_list) < node_mapping_length:
+            hard_neg_list.append(-1)
+        return hard_neg_list
+
+    pad_value_udf = F.udf(pad_mapped_values, ArrayType(IntegerType()))
+    # Make sure it keeps the original order
+    transformed_df = transformed_df.orderBy(ORDER_INDEX)
+    transformed_df = transformed_df.select(pad_value_udf(F.col(cols[0])).alias(cols[0]))
+
+    return transformed_df
+
+
+class DistHardEdgeNegativeTransformation(DistributedTransformation):
+    """Transformation to apply hard negative transformation.
+
+    Parameters
+    ----------
+    separator: str, optional
+        The separator for string input value. Only required when input value type is string.
+    spark: SparkSession
+        The spark session.
+    hard_node_mapping_dict: dict
+        The mapping dictionary contain mapping file directory and edge type.
+    """
+
+    def __init__(
+        self,
+        cols: Sequence[str],
+        spark: SparkSession,
+        separator: str = "",
+        hard_node_mapping_dict=None,
+    ) -> None:
+        super().__init__(cols, spark)
+        self.cols = cols
+        assert len(self.cols) == 1, "Hard Negative Transformation only supports single column"
+        self.separator = separator
+        self.hard_node_mapping_dict = hard_node_mapping_dict
+        assert self.hard_node_mapping_dict, "edge mapping dict cannot be None for hard negative "
+
+    def apply(self, input_df: DataFrame) -> DataFrame:
+        assert self.spark
+        transformed_df = apply_transform(
+            self.cols, self.separator, self.spark, input_df, self.hard_node_mapping_dict
+        )
+
+        return transformed_df
+
+    @staticmethod
+    def get_transformation_name() -> str:
+        return "DistHardEdgeNegativeTransformation"

graphstorm-processing/graphstorm_processing/graph_loaders/dist_heterogeneous_loader.py

@@ -1654,6 +1654,15 @@ def _process_edge_features(
                 .get(edge_type, {})
                 .get(feat_conf.feat_name, {})
             )
+            # Hard Negative Transformation use case, but should be able to be reused
+            if feat_conf.feat_type == "edge_dst_hard_negative":
+                hard_node_mapping_dict = {
+                    "edge_type": edge_type,
+                    "mapping_path": f"{self.output_prefix}/raw_id_mappings/",
+                    "format_name": FORMAT_NAME,
+                }
+                feat_conf.transformation_kwargs["hard_node_mapping_dict"] = hard_node_mapping_dict
+
             transformer = DistFeatureTransformer(feat_conf, self.spark, json_representation)
 
             if json_representation:

graphstorm-processing/tests/test_converter.py

@@ -401,7 +401,14 @@ def test_convert_gsprocessing(converter: GConstructConfigConverter):
             "files": ["/tmp/acm_raw/edges/author_writing_paper.parquet"],
             "source_id_col": "~from",
             "dest_id_col": "~to",
-            "features": [{"feature_col": ["author"], "feature_name": "feat"}],
+            "features": [
+                {"feature_col": ["author"], "feature_name": "feat"},
+                {
+                    "feature_col": ["author"],
+                    "feature_name": "hard_negative",
+                    "transform": {"name": "edge_dst_hard_negative", "separator": ";"},
+                },
+            ],
             "labels": [
                 {
                     "label_col": "edge_col",
@@ -505,7 +512,12 @@ def test_convert_gsprocessing(converter: GConstructConfigConverter):
     assert edges_output["dest"] == {"column": "~to", "type": "paper"}
     assert edges_output["relation"] == {"type": "writing"}
     assert edges_output["features"] == [
-        {"column": "author", "transformation": {"name": "no-op"}, "name": "feat"}
+        {"column": "author", "transformation": {"name": "no-op"}, "name": "feat"},
+        {
+            "column": "author",
+            "name": "hard_negative",
+            "transformation": {"name": "edge_dst_hard_negative", "kwargs": {"separator": ";"}},
+        },
     ]
     assert edges_output["labels"] == [
         {