Descriptive statistics on values outside of group for each group - apache-spark

Question

I have a Spark DataFrame like this:
edited: each name can appear multiple times, in any org.
df = sqlContext.createDataFrame(
[
('org_1', 'a', 1),
('org_1', 'a', 2),
('org_1', 'a', 3),
('org_1', 'b', 4),
('org_1', 'c', 5),
('org_2', 'a', 7),
('org_2', 'd', 4),
('org_2', 'e', 5),
('org_2', 'e', 10)
],
["org", "name", "value"]
)
I would like to calculate for each org and name: the mean, stddev and count of values from the rest of the names excluding that name within each org. E.g. For org_1, name b, mean = (1+2+3+5)/4
The DataFrame has ~450 million rows. I cannot use vectorized pandas_UDF because my Spark version is 2.2. There is also a constraint of spark.driver.maxResultSize of 4.0 GB.
I tried this on Pandas (filter rows within groups and take mean/std/count) on a DataFrame with only two columns (name and value). I haven't figured out how to do this with two levels of grouped columns (org and name).
def stats_fun(x):
return pd.Series({'data_mean': x['value'].mean(),
'data_std': x['value'].std(),
'data_n': x['value'].count(),
'anti_grp_mean': df[df['name'] != x.name]['value'].mean(),
'anti_grp_std': df[df['name'] != x.name]['value'].std(),
'anti_grp_n': df[df['name'] != x.name]['value'].count()
})
df.groupby('name').apply(stats_fun)
Is there a similar UDF function I can define on Spark? (This function would have to take in multiple columns). Otherwise, what is a more efficient way to do this?

Answer 1

A simple UDF can also work.
import pyspark.sql.functions as F
import numpy as np
from pyspark.sql.types import *
df = sql.createDataFrame(
[
('org_1', 'a', 1),
('org_1', 'a', 2),
('org_1', 'a', 3),
('org_1', 'b', 4),
('org_1', 'c', 5),
('org_2', 'a', 7),
('org_2', 'd', 4),
('org_2', 'e', 5),
('org_2', 'e', 10)
],
["org", "name", "value"]
)
+-----+----+-----+
| org|name|value|
+-----+----+-----+
|org_1| a| 1|
|org_1| a| 2|
|org_1| a| 3|
|org_1| b| 4|
|org_1| c| 5|
|org_2| a| 7|
|org_2| d| 4|
|org_2| e| 5|
|org_2| e| 10|
+-----+----+-----+
After applying groupby and collecting all elements in list, we apply udf to find statistics. After that, columns are exploded and split into multiple columns.
def _find_stats(a,b):
dict_ = zip(a,b)
stats = []
for name in a:
to_cal = [v for k,v in dict_ if k != name]
stats.append((name,float(np.mean(to_cal))\
,float(np.std(to_cal))\
,len(to_cal)))
print stats
return stats
find_stats = F.udf(_find_stats,ArrayType(ArrayType(StringType())))
cols = ['name', 'mean', 'stddev', 'count']
splits = [F.udf(lambda val:val[0],StringType()),\
F.udf(lambda val:val[1],StringType()),\
F.udf(lambda val:val[2],StringType()),\
F.udf(lambda val:val[3],StringType())]
df = df.groupby('org').agg(*[F.collect_list('name').alias('name'), F.collect_list('value').alias('value')])\
.withColumn('statistics', find_stats(F.col('name'), F.col('value')))\
.drop('name').drop('value')\
.select('org', F.explode('statistics').alias('statistics'))\
.select(['org']+[split_('statistics').alias(col_name) for split_,col_name in zip(splits,cols)])\
.dropDuplicates()
df.show()
+-----+----+-----------------+------------------+-----+
| org|name| mean| stddev|count|
+-----+----+-----------------+------------------+-----+
|org_1| c| 2.5| 1.118033988749895| 4|
|org_2| e| 5.5| 1.5| 2|
|org_2| a|6.333333333333333|2.6246692913372702| 3|
|org_2| d|7.333333333333333|2.0548046676563256| 3|
|org_1| a| 4.5| 0.5| 2|
|org_1| b| 2.75| 1.479019945774904| 4|
+-----+----+-----------------+------------------+-----+
If you also want 'value' column, you can insert that in the tuple in udf function and add one split udf.
Also, since there will be duplicates in the dataframe due to repetition of names, you can remove them using dropDuplicates.

Answer 2

Here is a way to do this using only DataFrame functions.
Just join your DataFrame to itself on the org column and use a where clause to specify that the name column should be different. Then we select the distinct rows of ('l.org', 'l.name', 'r.name', 'r.value') - essentially, we ignore the l.value column because we want to avoid double counting for the same (org, name) pair.
For example, this is how you could collect the other values for each ('org', 'name') pair:
import pyspark.sql.functions as f
df.alias('l').join(df.alias('r'), on='org')\
.where('l.name != r.name')\
.select('l.org', 'l.name', 'r.name', 'r.value')\
.distinct()\
.groupBy('l.org', 'l.name')\
.agg(f.collect_list('r.value').alias('other_values'))\
.show()
#+-----+----+------------+
#| org|name|other_values|
#+-----+----+------------+
#|org_1| a| [4, 5]|
#|org_1| b|[1, 2, 3, 5]|
#|org_1| c|[1, 2, 3, 4]|
#|org_2| a| [4, 5, 10]|
#|org_2| d| [7, 5, 10]|
#|org_2| e| [7, 4]|
#+-----+----+------------+
For the descriptive stats, you can use the mean, stddev, and count functions from pyspark.sql.functions:
df.alias('l').join(df.alias('r'), on='org')\
.where('l.name != r.name')\
.select('l.org', 'l.name', 'r.name', 'r.value')\
.distinct()\
.groupBy('l.org', 'l.name')\
.agg(
f.mean('r.value').alias('mean'),
f.stddev('r.value').alias('stddev'),
f.count('r.value').alias('count')
)\
.show()
#+-----+----+-----------------+------------------+-----+
#| org|name| mean| stddev|count|
#+-----+----+-----------------+------------------+-----+
#|org_1| a| 4.5|0.7071067811865476| 2|
#|org_1| b| 2.75| 1.707825127659933| 4|
#|org_1| c| 2.5|1.2909944487358056| 4|
#|org_2| a|6.333333333333333|3.2145502536643185| 3|
#|org_2| d|7.333333333333333|2.5166114784235836| 3|
#|org_2| e| 5.5|2.1213203435596424| 2|
#+-----+----+-----------------+------------------+-----+
Note that pyspark.sql.functions.stddev() returns the unbiased sample standard deviation. If you wanted the population standard deviation, use pyspark.sql.functions.stddev_pop():
df.alias('l').join(df.alias('r'), on='org')\
.where('l.name != r.name')\
.groupBy('l.org', 'l.name')\
.agg(
f.mean('r.value').alias('mean'),
f.stddev_pop('r.value').alias('stddev'),
f.count('r.value').alias('count')
)\
.show()
#+-----+----+-----------------+------------------+-----+
#| org|name| mean| stddev|count|
#+-----+----+-----------------+------------------+-----+
#|org_1| a| 4.5| 0.5| 2|
#|org_1| b| 2.75| 1.479019945774904| 4|
#|org_1| c| 2.5| 1.118033988749895| 4|
#|org_2| a|6.333333333333333|2.6246692913372702| 3|
#|org_2| d|7.333333333333333|2.0548046676563256| 3|
#|org_2| e| 5.5| 1.5| 2|
#+-----+----+-----------------+------------------+-----+
EDIT
As #NaomiHuang mentioned in the comments, you could also reduce l to the distinct org/name pairs before doing the join:
df.select('org', 'name')\
.distinct()\
.alias('l')\
.join(df.alias('r'), on='org')\
.where('l.name != r.name')\
.groupBy('l.org', 'l.name')\
.agg(f.collect_list('r.value').alias('other_values'))\
.show()
#+-----+----+------------+
#| org|name|other_values|
#+-----+----+------------+
#|org_1| a| [5, 4]|
#|org_1| b|[5, 1, 2, 3]|
#|org_1| c|[1, 2, 3, 4]|
#|org_2| a| [4, 5, 10]|
#|org_2| d| [7, 5, 10]|
#|org_2| e| [7, 4]|
#+-----+----+------------+