Count occurrences of a list of substrings in a pyspark df column

Question

I want to count the occurrences of list of substrings and create a column based on a column in the pyspark df which contains a long string.

Input:          
       ID    History

       1     USA|UK|IND|DEN|MAL|SWE|AUS
       2     USA|UK|PAK|NOR
       3     NOR|NZE
       4     IND|PAK|NOR

 lst=['USA','IND','DEN']


Output :
       ID    History                      Count

       1     USA|UK|IND|DEN|MAL|SWE|AUS    3
       2     USA|UK|PAK|NOR                1
       3     NOR|NZE                       0
       4     IND|PAK|NOR                   1

Answer 1

# Importing requisite packages and creating a DataFrame
from pyspark.sql.functions import split, col, size, regexp_replace
values = [(1,'USA|UK|IND|DEN|MAL|SWE|AUS'),(2,'USA|UK|PAK|NOR'),(3,'NOR|NZE'),(4,'IND|PAK|NOR')]
df = sqlContext.createDataFrame(values,['ID','History'])
df.show(truncate=False)
+---+--------------------------+
|ID |History                   |
+---+--------------------------+
|1  |USA|UK|IND|DEN|MAL|SWE|AUS|
|2  |USA|UK|PAK|NOR            |
|3  |NOR|NZE                   |
|4  |IND|PAK|NOR               |
+---+--------------------------+

这个想法是根据这三个delimiters:来分割字符串lst=['USA','IND','DEN']，然后计算生成的子字符串的数量。

例如；字符串USA|UK|IND|DEN|MAL|SWE|AUS被分割为 - ,, |UK|, |, |MAL|SWE|AUS。由于创建了 4 个子字符串并且有 3 个分隔符匹配，因此4-1 = 3给出了出现在列字符串中的这些字符串的计数。

我不确定 Spark 中是否支持多字符分隔符，因此第一步，我们将列表中的这 3 个子字符串中的任何一个替换['USA','IND','DEN']为 flag/dummy value %。您也可以使用其他东西。以下代码执行此操作replacement-

df = df.withColumn('History_X',col('History'))
lst=['USA','IND','DEN']
for i in lst:
    df = df.withColumn('History_X', regexp_replace(col('History_X'), i, '%'))
df.show(truncate=False)
+---+--------------------------+--------------------+
|ID |History                   |History_X           |
+---+--------------------------+--------------------+
|1  |USA|UK|IND|DEN|MAL|SWE|AUS|%|UK|%|%|MAL|SWE|AUS|
|2  |USA|UK|PAK|NOR            |%|UK|PAK|NOR        |
|3  |NOR|NZE                   |NOR|NZE             |
|4  |IND|PAK|NOR               |%|PAK|NOR           |
+---+--------------------------+--------------------+

splitting最后，我们先计算它作为分隔符创建的子字符串的数量%，然后计算使用 function 创建的子字符串的数量size，最后减去 1。

df = df.withColumn('Count', size(split(col('History_X'), "%")) - 1).drop('History_X')
df.show(truncate=False)
+---+--------------------------+-----+
|ID |History                   |Count|
+---+--------------------------+-----+
|1  |USA|UK|IND|DEN|MAL|SWE|AUS|3    |
|2  |USA|UK|PAK|NOR            |1    |
|3  |NOR|NZE                   |0    |
|4  |IND|PAK|NOR               |1    |
+---+--------------------------+-----+

Answer 2

如果您使用Spark 2.4+，您可以尝试 SPARK SQL 高阶函数filter()：

from pyspark.sql import functions as F

>>> df.show(5,0)
+---+--------------------------+
|ID |History                   |
+---+--------------------------+
|1  |USA|UK|IND|DEN|MAL|SWE|AUS|
|2  |USA|UK|PAK|NOR            |
|3  |NOR|NZE                   |
|4  |IND|PAK|NOR               |
+---+--------------------------+

df_new = df.withColumn('data', F.split('History', '\|')) \
           .withColumn('cnt', F.expr('size(filter(data, x -> x in ("USA", "IND", "DEN")))'))

>>> df_new.show(5,0)
+---+--------------------------+----------------------------------+---+
|ID |History                   |data                              |cnt|
+---+--------------------------+----------------------------------+---+
|1  |USA|UK|IND|DEN|MAL|SWE|AUS|[USA, UK, IND, DEN, MAL, SWE, AUS]|3  |
|2  |USA|UK|PAK|NOR            |[USA, UK, PAK, NOR]               |1  |
|3  |NOR|NZE                   |[NOR, NZE]                        |0  |
|4  |IND|PAK|NOR               |[IND, PAK, NOR]                   |1  |
+---+--------------------------+----------------------------------+---+

我们首先将字段拆分History为名为的数组列data，然后使用过滤器函数：

filter(data, x -> x in ("USA", "IND", "DEN"))

仅检索满足条件的数组元素：IN ("USA", "IND", "DEN")，之后，我们使用size()函数对结果数组进行计数。

更新：添加了另一种使用array_contains()的方法，该方法适用于旧版本 Spark：

lst = ["USA", "IND", "DEN"]

df_new = df.withColumn('data', F.split('History', '\|')) \
           .withColumn('Count', sum([F.when(F.array_contains('data',e),1).otherwise(0) for e in lst]))

注意：数组中的重复条目将被跳过，此方法仅计算唯一的国家/地区代码。