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               |
+---+--------------------------+

The idea is to split the string based on these three delimiters: lst=['USA','IND','DEN'] and then count the number of substrings produced.

For eg; the string USA|UK|IND|DEN|MAL|SWE|AUS gets split like - ,, |UK|, |, |MAL|SWE|AUS. Since, there were 4 substrings created and there were 3 delimiters matches, so 4-1 = 3 gives the count of these strings appearing in the column string.

I am not sure if multi character delimiters are supported in Spark, so as a first step, we replace any of these 3 sub-strings in the list ['USA','IND','DEN'] with a flag/dummy value %. You could use something else as well. The following code does this 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           |
+---+--------------------------+--------------------+

Finally, we count the number of substrings created by splitting it first with % being the delimiter, then counting the number of substrings created with size function and finally subtracting 1 from it.

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

If you are using Spark 2.4+, you can try the SPARK SQL higher order function 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  |
+---+--------------------------+----------------------------------+---+

Where we first split the field History into an array column called data and then use the filter function:

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

to retrieve only array elements which satisfy the condition: IN ("USA", "IND", "DEN"), after that, we count the resulting array with size() function.

UPDATE: Added another way to use array_contains() which should works for old version 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]))

Note: duplicate entries in arrays will be skipped, this method only counts unique Country code.