Merge pull request #19 from SCAI-BIO/fix/corr-nan

Fix: add test

syndat/scores.py

@@ -203,6 +203,12 @@ def correlation(real: pd.DataFrame, synthetic: pd.DataFrame, score=True) -> floa
     # Compute numerical correlation only
     real_numerical = real_encoded.select_dtypes(include=[np.number])
     synthetic_numerical = synthetic_encoded.select_dtypes(include=[np.number])
+    # Remove constant columns (zero variance)
+    constant_columns = real_numerical.columns[real_numerical.nunique() <= 1]
+    if len(constant_columns) > 0:
+        logger.warning(f'Removing constant columns {constant_columns} for correlation computation.')
+        real_numerical = real_numerical.drop(columns=constant_columns, errors="ignore")
+        synthetic_numerical = synthetic_numerical.drop(columns=constant_columns, errors="ignore")
     # Compute correlation matrices
     corr_real = real_numerical.corr()
     corr_synthetic = synthetic_numerical.corr()
@@ -215,7 +221,8 @@ def correlation(real: pd.DataFrame, synthetic: pd.DataFrame, score=True) -> floa
         corr_real = corr_real.drop(columns=one_hot_encoded_columns)
         corr_synthetic = corr_synthetic.drop(columns=one_hot_encoded_columns)
     # now compute correlation matrices
-    norm_diff = np.linalg.norm(corr_real - corr_synthetic)
+    corr_diff = corr_real - corr_synthetic
+    norm_diff = np.linalg.norm(corr_diff)
     norm_real = np.linalg.norm(corr_real)
     norm_quotient = norm_diff / norm_real
     if score:

tests/test_correlation.py

@@ -116,3 +116,16 @@ def test_correlation_with_only_categorical(self):
         # Adjust the expected result based on encoding and correlation of categorical data
         self.assertLess(result, 100, "Correlation score should be less than 100 for datasets with different "
                                      "categorical data")
+
+    def test_correlation_constant_column(self):
+        real_data = pd.DataFrame({
+            'A': [1, 1, 1, 1, 1],
+            'B': [2, 3, 4, 5, 6]
+        })
+        synthetic_data = pd.DataFrame({
+            'A': [1, 1, 1, 1, 1],
+            'B': [2, 3, 4, 5, 6]
+        })
+        result = correlation(real_data, synthetic_data)
+        # Depending on the implementation, the correlation might return NaN or handle it explicitly
+        self.assertFalse(pd.isna(result), "Correlation score should not result in NaN even with a constant column")