stabilize normalization results by modifying sampling process

featuretools/mkfeat/normalize.py

@@ -2,9 +2,20 @@
 import autonormalize as an
 
 
-def normalize(df: DataFrame, key_colname):
+def sampling(df: DataFrame):
     if len(df) > 1000:
-        df = df.sample(n=1000)
+        ret = df
+        repeat = len(df) // 1000 if len(df) // 1000 > 5 else 5
+        for _ in range(repeat):
+            sample = df.sample(n=1000)
+            ret = ret & sample
+        return ret
+    else:
+        return df
+
+
+def normalize(df: DataFrame, key_colname):
+    df = sampling(df)
     es = an.auto_entityset(df, index=key_colname, accuracy=0.98)
 
     norminfos = []