Filter out keys that cannot be statistically modeled (#525)

sdmetrics/single_table/detection/base.py

@@ -9,7 +9,7 @@
 from sdmetrics.errors import IncomputableMetricError
 from sdmetrics.goal import Goal
 from sdmetrics.single_table.base import SingleTableMetric
-from sdmetrics.utils import HyperTransformer
+from sdmetrics.utils import HyperTransformer, get_alternate_keys
 
 LOGGER = logging.getLogger(__name__)
 
@@ -43,6 +43,32 @@ def _fit_predict(X_train, y_train, X_test):
         """Fit a classifier and then use it to predict."""
         raise NotImplementedError()
 
+    @staticmethod
+    def _drop_non_compute_columns(real_data, synthetic_data, metadata):
+        """Drop all columns that cannot be statistically modeled."""
+        transformed_real_data = real_data
+        transformed_synthetic_data = synthetic_data
+
+        if metadata is not None:
+            drop_columns = []
+            drop_columns.extend(get_alternate_keys(metadata))
+            for column in metadata.get('columns', []):
+                if ('primary_key' in metadata and
+                        (column == metadata['primary_key'] or
+                         column in metadata['primary_key'])):
+                    drop_columns.append(column)
+
+                column_info = metadata['columns'].get(column, {})
+                sdtype = column_info.get('sdtype')
+                pii = column_info.get('pii')
+                if sdtype not in ['numerical', 'datetime', 'categorical'] or pii:
+                    drop_columns.append(column)
+
+            if drop_columns:
+                transformed_real_data = real_data.drop(drop_columns, axis=1)
+                transformed_synthetic_data = synthetic_data.drop(drop_columns, axis=1)
+        return transformed_real_data, transformed_synthetic_data
+
     @classmethod
     def compute(cls, real_data, synthetic_data, metadata=None):
         """Compute this metric.
@@ -68,13 +94,8 @@ def compute(cls, real_data, synthetic_data, metadata=None):
         real_data, synthetic_data, metadata = cls._validate_inputs(
             real_data, synthetic_data, metadata)
 
-        if metadata is not None and 'primary_key' in metadata:
-            transformed_real_data = real_data.drop(metadata['primary_key'], axis=1)
-            transformed_synthetic_data = synthetic_data.drop(metadata['primary_key'], axis=1)
-
-        else:
-            transformed_real_data = real_data
-            transformed_synthetic_data = synthetic_data
+        transformed_real_data, transformed_synthetic_data = cls._drop_non_compute_columns(
+            real_data, synthetic_data, metadata)
 
         ht = HyperTransformer()
         transformed_real_data = ht.fit_transform(transformed_real_data).to_numpy()

tests/unit/single_table/detection/test_detection.py

@@ -65,3 +65,80 @@ def test_primary_key_detection_metrics(self, fit_transform_mock, transform_mock)
         transform_mock.assert_called_with(expected_return_synthetic)
         assert expected_return_real == call_1
         assert expected_return_synthetic == call_2
+
+    @patch('sdmetrics.utils.HyperTransformer.transform')
+    @patch('sdmetrics.utils.HyperTransformer.fit_transform')
+    def test_ignore_keys_detection_metrics(self, fit_transform_mock, transform_mock):
+        """This test checks that ``primary_key`` columns of dataset are ignored.
+
+        Ensure that ``primary_keys`` are ignored for Detection metrics expect that the match
+        is made correctly.
+        """
+
+        # Setup
+        real_data = pd.DataFrame({
+            'ID_1': [1, 2, 1, 3, 4],
+            'col1': [43.0, 47.5, 34.2, 30.3, 39.1],
+            'col2': [1.0, 2.0, 3.0, 4.0, 5.0],
+            'ID_2': ['aa', 'bb', 'cc', 'dd', 'bb'],
+            'col3': [5, 6, 7, 8, 9],
+            'ID_3': ['a', 'b', 'c', 'd', 'e'],
+            'blob': ['Hello world!', 'Hello world!', 'This is SDV', 'This is SDV', 'Hello world!'],
+            'col4': [1, 3, 9, 2, 1],
+            'col5': [10, 20, 30, 40, 50]
+        })
+        synthetic_data = pd.DataFrame({
+            'ID_1': [1, 3, 4, 2, 2],
+            'col1': [23.0, 47.1, 44.9, 31.3, 9.7],
+            'col2': [11.0, 22.0, 33.0, 44.0, 55.0],
+            'ID_2': ['aa', 'bb', 'cc', 'dd', 'ee'],
+            'col3': [55, 66, 77, 88, 99],
+            'ID_3': ['a', 'b', 'e', 'd', 'c'],
+            'blob': ['Hello world!', 'Hello world!', 'This is SDV', 'This is SDV', 'Hello world!'],
+            'col4': [4, 1, 3, 1, 9],
+            'col5': [10, 20, 30, 40, 50]
+        })
+        metadata = {
+            'columns': {
+                'ID_1': {'sdtype': 'numerical'},
+                'col1': {'sdtype': 'numerical', 'pii': True},
+                'col2': {'sdtype': 'numerical'},
+                'ID_2': {'sdtype': 'categorical'},
+                'col3': {'sdtype': 'numerical'},
+                'ID_3': {'sdtype': 'id'},
+                'blob': {'sdtype': 'text'},
+                'col4': {'sdtype': 'numerical', 'pii': False},
+                'col5': {'sdtype': 'numerical'}
+            },
+            'primary_key': {'ID_1', 'ID_2'},
+            'alternate_keys': ['col5']
+        }
+
+        expected_real_dataframe = pd.DataFrame({
+            'col2': [1.0, 2.0, 3.0, 4.0, 5.0],
+            'col3': [5, 6, 7, 8, 9],
+            'col4': [1, 3, 9, 2, 1]
+        })
+        expected_synthetic_dataframe = pd.DataFrame({
+            'col2': [11.0, 22.0, 33.0, 44.0, 55.0],
+            'col3': [55, 66, 77, 88, 99],
+            'col4': [4, 1, 3, 1, 9]
+        })
+
+        expected_return_real = DataFrameMatcher(expected_real_dataframe)
+        expected_return_synthetic = DataFrameMatcher(expected_synthetic_dataframe)
+        fit_transform_mock.return_value = expected_real_dataframe
+        transform_mock.return_value = expected_synthetic_dataframe
+
+        # Run
+        LogisticDetection().compute(real_data, synthetic_data, metadata)
+
+        # Assert
+
+        # check that ``fit_transform`` and  ``transform`` received the good argument.
+        call_1 = pd.DataFrame(fit_transform_mock.call_args_list[0][0][0])
+        call_2 = pd.DataFrame(transform_mock.call_args_list[0][0][0])
+
+        transform_mock.assert_called_with(expected_return_synthetic)
+        assert expected_return_real == call_1
+        assert expected_return_synthetic == call_2