Merge pull request #47 from outbrain/some-fixes

Renormalization for randomized-based heuristics
outbrain-inc · Oct 9, 2023 · 2fc30f2 · 2fc30f2
2 parents 05f9039 + 1c8b7b7
commit 2fc30f2
Show file tree

Hide file tree

Showing 3 changed files with 10 additions and 9 deletions.
diff --git a/outrank/algorithms/feature_ranking/ranking_mi_numba.py b/outrank/algorithms/feature_ranking/ranking_mi_numba.py
@@ -81,24 +81,25 @@ def compute_entropies(
 
         initial_prob = _f_value_counts / all_events
         x_value_subspace = np.where(X == f_values[f_index])
+
         Y_classes = Y[x_value_subspace]
-        index = 0
+        Y_classes_spoofed = np.roll(Y, _f_value_counts)[x_value_subspace]
+
         nonzero_class_counts = np.zeros(len(class_values), dtype=np.int32)
+        nonzero_class_counts_spoofed = np.zeros(len(class_values), dtype=np.int32)
 
         # Cache nonzero counts
-        for c in class_values:
+        for index, c in enumerate(class_values):
             nonzero_class_counts[index] = np.count_nonzero(Y_classes == c)
-            index += 1
+            nonzero_class_counts_spoofed[index] = np.count_nonzero(Y_classes_spoofed == c)
+
         conditional_entropy += compute_conditional_entropy(
             Y_classes, class_values, _f_value_counts, initial_prob, nonzero_class_counts,
         )
 
         if cardinality_correction:
-            # A neat hack that seems to work fine (permutations are expensive)
-            Y_classes = np.roll(Y, _f_value_counts)[x_value_subspace]
-
             background_cond_entropy += compute_conditional_entropy(
-                Y_classes, class_values, _f_value_counts, initial_prob, nonzero_class_counts,
+                Y_classes_spoofed, class_values, _f_value_counts, initial_prob, nonzero_class_counts_spoofed,
             )
 
     if not cardinality_correction:

diff --git a/outrank/core_ranking.py b/outrank/core_ranking.py
@@ -72,9 +72,9 @@ def mixed_rank_graph(
 
     # Handle cont. types prior to interaction evaluation
     pbar.set_description('Encoding columns')
-    col_dots = '.'
     start_enc_timer = timer()
     tmp_df = pd.DataFrame({k : tmp_df[k].cat.codes for k in all_columns})
+
     end_enc_timer = timer()
     out_time_struct['encoding_columns'] = end_enc_timer - start_enc_timer
 

diff --git a/setup.py b/setup.py
@@ -23,7 +23,7 @@ def _read_description():
 packages = [x for x in setuptools.find_packages() if x != 'test']
 setuptools.setup(
     name='outrank',
-    version='0.94.2',
+    version='0.95',
     description='OutRank: Feature ranking for massive sparse data sets.',
     long_description=_read_description(),
     long_description_content_type='text/markdown',