计算 Spark DataFrame 中分组数据的标准差 [英] Calculate the standard deviation of grouped data in a Spark DataFrame

问题描述

我有从 csv 中获取并转换为 DataFrame 的用户日志，以便利用 SparkSQL 查询功能.单个用户每小时会创建很多条目，我想为每个用户收集一些基本的统计信息；实际上只是用户实例的计数、平均值和众多列的标准偏差.我能够通过使用 groupBy($"user") 和带有 SparkSQL 函数的聚合器来快速获取平均值和计数信息:

I have user logs that I have taken from a csv and converted into a DataFrame in order to leverage the SparkSQL querying features. A single user will create numerous entries per hour, and I would like to gather some basic statistical information for each user; really just the count of the user instances, the average, and the standard deviation of numerous columns. I was able to quickly get the mean and count information by using groupBy($"user") and the aggregator with SparkSQL functions for count and avg:

val meanData = selectedData.groupBy($"user").agg(count($"logOn"),
avg($"transaction"), avg($"submit"), avg($"submitsPerHour"), avg($"replies"),
avg($"repliesPerHour"), avg($"duration"))

但是，我似乎找不到一种同样优雅的方法来计算标准偏差.到目前为止，我只能通过映射字符串、双对并使用 StatCounter().stdev 实用程序来计算它:

However, I cannot seem to find an equally elegant way to calculate the standard deviation. So far I can only calculate it by mapping a string, double pair and use StatCounter().stdev utility:

val stdevduration = duration.groupByKey().mapValues(value =>
org.apache.spark.util.StatCounter(value).stdev)

然而，这会返回一个 RDD，我想尝试将其全部保存在 DataFrame 中，以便对返回的数据进行进一步查询.

This returns an RDD however, and I would like to try and keep it all in a DataFrame for further queries to be possible on the returned data.

推荐答案

Spark 1.6+

您可以使用stddev_pop 计算总体标准差，使用stddev/stddev_samp 计算无偏样本标准差:

You can use stddev_pop to compute population standard deviation and stddev / stddev_samp to compute unbiased sample standard deviation:

import org.apache.spark.sql.functions.{stddev_samp, stddev_pop}

selectedData.groupBy($"user").agg(stdev_pop($"duration"))

Spark 1.5 及以下(原始答案):

不是那么漂亮和有偏见(与从 describe 返回的值相同)但使用公式:

Not so pretty and biased (same as the value returned from describe) but using formula:

你可以这样做:

import org.apache.spark.sql.functions.sqrt

selectedData
    .groupBy($"user")
    .agg((sqrt(
        avg($"duration" * $"duration") -
        avg($"duration") * avg($"duration")
     )).alias("duration_sd"))

你当然可以创建一个函数来减少混乱:

You can of course create a function to reduce the clutter:

import org.apache.spark.sql.Column
def mySd(col: Column): Column = {
    sqrt(avg(col * col) - avg(col) * avg(col))
}

df.groupBy($"user").agg(mySd($"duration").alias("duration_sd"))

也可以使用 Hive UDF:

It is also possible to use Hive UDF:

df.registerTempTable("df")
sqlContext.sql("""SELECT user, stddev(duration)
                  FROM df
                  GROUP BY user""")

<小时>

图片来源:https://en.wikipedia.org/wiki/Standard_deviation

这篇关于计算 Spark DataFrame 中分组数据的标准差的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！