Spark DataFrame: Computing row-wise mean (or any a

问题:

I have a Spark DataFrame loaded up in memory, and I want to take the mean (or any aggregate operation) over the columns. How would I do that? (In numpy, this is known as taking an operation over axis=1).

If one were calculating the mean of the DataFrame down the rows (axis=0), then this is already built in:

from pyspark.sql import functions as F
F.mean(...)

But is there a way to programmatically do this against the entries in the columns? For example, from the DataFrame below

+--+--+---+---+
|id|US| UK|Can|
+--+--+---+---+
| 1|50|  0|  0|
| 1| 0|100|  0|
| 1| 0|  0|125|
| 2|75|  0|  0|
+--+--+---+---+

Omitting id, the means would be

+------+
|  mean|
+------+
| 16.66|
| 33.33|
| 41.67|
| 25.00|
+------+

回答1:

All you need here is a standard SQL like this:

SELECT (US + UK + CAN) / 3 AS mean FROM df

which can be used directly with SqlContext.sql or expressed using DSL

df.select(((col("UK") + col("US") + col("CAN")) / lit(3)).alias("mean"))

If you have a larger number of columns you can generate expression as follows:

from functools import reduce
from operator import add
from pyspark.sql.functions import col, lit

n = lit(len(df.columns) - 1.0)
rowMean  = (reduce(add, (col(x) for x in df.columns[1:])) / n).alias("mean")

df.select(rowMean)

or

rowMean  = (sum(col(x) for x in df.columns[1:]) / n).alias("mean")
df.select(rowMean)

Finally its equivalent in Scala:

df.select(df.columns
  .drop(1)
  .map(col)
  .reduce(_ + _)
  .divide(df.columns.size - 1)
  .alias("mean"))

In a more complex scenario you can combine columns using array function and use an UDF to compute statistics:

import numpy as np
from pyspark.sql.functions import array, udf
from pyspark.sql.types import FloatType

combined = array(*(col(x) for x in df.columns[1:]))
median_udf = udf(lambda xs: float(np.median(xs)), FloatType())

df.select(median_udf(combined).alias("median"))

The same operation expressed using Scala API:

val combined = array(df.columns.drop(1).map(col).map(_.cast(DoubleType)): _*)
val median_udf = udf((xs: Seq[Double]) => 
    breeze.stats.DescriptiveStats.percentile(xs, 0.5))

df.select(median_udf(combined).alias("median"))

Since Spark 2.4 an alternative approach is to combine values into an array and apply aggregate expression. See for example Spark Scala row-wise average by handling null.

回答2:

in Scala something like this would do it

val cols = Seq("US","UK","Can")
f.map(r => (r.getAs[Int]("id"),r.getValuesMap(cols).values.fold(0.0)(_+_)/cols.length)).toDF