stabilize normalization results by modifying sampling process

featuretools/mkfeat/normalize.py

@@ -1,33 +1,41 @@
 from pandas import DataFrame
 import autonormalize as an
 
-
 def normalize(df: DataFrame, key_colname):
+    es = None
+    entities = set()
+    relationships = set()
+
     if len(df) > 1000:
-        df = df.sample(n=1000)
-    es = an.auto_entityset(df, index=key_colname, accuracy=0.98)
+        for _ in range(5):  
+            df = df.sample(n=1000)
+            es = an.auto_entityset(df, index=key_colname, accuracy=0.98)
+            entities.update(es.entities[1:])
+            relationships.update(es.relationships)
+    else:
+        es = an.auto_entityset(df, index=key_colname, accuracy=0.98)
+        entities.update(es.entities[1:])
+        relationships.update(es.relationships)
 
     norminfos = []
     # 첫번째 이외의 entity들에 대해서. 첫번째 entity가 main임을 가정
-    entities = es.entities[1:]
     for et in entities:
         norminfo = []
         for var in et.variables:
             norminfo.append(var.name)
         norminfos.append(norminfo)
     for norminfo in norminfos:
-        parent_ids = _get_parent_entity_ids(es, norminfo[0])
+        parent_ids = _get_parent_entity_ids(relationships, norminfo[0])
         for parent_id in parent_ids:
             vars = es[parent_id].variables
             for var in vars[1:]:
                 norminfo.append(var.name)
     return norminfos
 
-
-def _get_parent_entity_ids(es, child_id):
+def _get_parent_entity_ids(rels, child_id):
     parent_ids = []
-    for rel in es.relationships:
+    for rel in rels:
         if child_id == rel.child_entity.id:
             parent_ids.append(rel.parent_entity.id)
-            parent_ids += _get_parent_entity_ids(es, rel.parent_entity.id)
-    return parent_ids
+            parent_ids += _get_parent_entity_ids(rels, rel.parent_entity.id)
+    return parent_ids