PySpark - 如何遍历数据帧并匹配另一个数据帧中的另一个常见值

Question

This is a recommender system and I have a Dataframe which contains about 10 recommended item for each user ( recommendation_df ) and I have another Dataframe which consist of the recent purchases of each user ( recent_df ).

I am trying to code out this task but I can't seem to get along the syntax, and the manipulation

I am implementing a hit/miss ratio, basically for every new_party_id in recent_df , if any of the merch_store_code matches the merch_store_code for the same party_id in the recommendation_df , count +=1 (Hit)

Then calculating the hit/miss ratio by using count/total user count

(However in recent_df, each user might have multiple recent purchases, but if any of the purchases is in the list of recommendations_list for the same user, take it as a hit (count +=1)

recommendation_df

+--------------+----------------+-----------+----------+
|party_id_index|merch_store_code|     rating|  party_id|
+--------------+----------------+-----------+----------+
|           148|       900000166|  0.4021678|G18B00332C|
|           148|       168339566| 0.27687865|G18B00332C|
|           148|       168993309| 0.15999989|G18B00332C|
|           148|       168350313|  0.1431974|G18B00332C|
|           148|       168329726| 0.13634883|G18B00332C|
|           148|       168351967|0.120235085|G18B00332C|
|           148|       168993312| 0.11800903|G18B00332C|
|           148|       168337234|0.116267696|G18B00332C|
|           148|       168993256| 0.10836013|G18B00332C|
|           148|       168339482| 0.10341005|G18B00332C|
|           463|       168350313| 0.93455887|K18M926299|
|           463|       900000072|  0.8275664|K18M926299|
|           463|       700012303| 0.70220494|K18M926299|
|           463|       700012180| 0.23209469|K18M926299|
|           463|       900000157|  0.1727839|K18M926299|
|           463|       700013689| 0.13854747|K18M926299|
|           463|       900000166| 0.12866624|K18M926299|
|           463|       168993284|0.107065596|K18M926299|
|           463|       168993269| 0.10272527|K18M926299|
|           463|       168339566| 0.10256036|K18M926299|
+--------------+----------------+-----------+----------+

recent_df

+------------+---------------+----------------+
|new_party_id|recent_purchase|merch_store_code|
+------------+---------------+----------------+
|  A11275842R|     2022-05-21|       168289403|
|  A131584211|     2022-06-01|       168993311|
|  A131584211|     2022-06-01|       168349493|
|  A131584211|     2022-06-01|       168350192|
|  A182P3539K|     2022-03-26|       168341707|
|  A182V2883F|     2022-05-26|       168350824|
|  A183B5482P|     2022-05-10|       168993464|
|  A183C6900K|     2022-05-14|       168338795|
|  A183D56093|     2022-05-20|       700012303|
|  A183J5388G|     2022-03-18|       700013650|
|  A183U8880P|     2022-04-01|       900000072|
|  A183U8880P|     2022-04-01|       168991904|
|  A18409762L|     2022-05-10|       168319352|
|  A18431276J|     2022-05-14|       168163905|
|  A18433684M|     2022-03-21|       168993324|
|  A18433978F|     2022-05-20|       168341876|
|  A184410389|     2022-05-04|       900000166|
|  A184716280|     2022-04-06|       700013653|
|  A18473797O|     2022-05-24|       168330339|
|  A18473797O|     2022-05-24|       168350592|
+------------+---------------+----------------+

Here is my current coding logic:

count = 0
def hitratio(recommendation_df, recent_df):
 for i in recent_df['new_party_id']:
  for j  in recommendation_df['party_id']:
   if (i = j) & i.merch_store_code == j.merch_store_code:
    count += 1
  return (count/recent_df.count())

Answer 1

assumption : i am taking all the count rows of recent df as denominator for calculating the hit/miss ratio you can change the formula.

from pyspark.sql import functions as F
matching_cond = ((recent_df["merch_store_code"]==recommender_df["merch_store_code"]) &(recommendation_df["party_id"].isNotNull()))

df_recent_fnl= df_recent.join(recommendation_df,df_recent["new_party_id"]==recommendation_df["party_id"],"left")\
.select(df_recent["*"],recommender_df["merch_store_code"],recommendation_df["party_id"])\
.withColumn("hit",F.when(matching_cond,F.lit(True)).otherwise(F.lit(False)))\
.withColumn("hit/miss",df_recent_fnl.filter(F.col("hit")).count()/df_recent.count())

do let me know if you have any questions around this .

If you like my solution , you can upvote

Answer 2

In Spark, refrain from loops on rows. Spark does not work like that, you need to think of the whole column, not about row-by-row scenario.

You need to join both tables and select users, but they need to be without duplicates (distinct)

from pyspark.sql import functions as F
df_distinct_matches = (
    recent_df
    .join(recommendation_df, F.col('new_party_id') == F.col('party_id'))
    .select('party_id').distinct()
)
hit = df_distinct_matches.count()