为 PySpark 中的最大值选择每行的列名

Question

我有一个这样的数据框，只显示了两列，但是原始数据框中有很多列

data = [(("ID1", 3, 5)), (("ID2", 4, 12)), (("ID3", 8, 3))]
df = spark.createDataFrame(data, ["ID", "colA", "colB"])
df.show()

+---+----+----+
| ID|colA|colB|
+---+----+----+
|ID1|   3|   5|
|ID2|   4|  12|
|ID3|   8|   3|
+---+----+----+

我想提取每行列的名称，它具有最大值。因此预期的输出是这样的

+---+----+----+-------+
| ID|colA|colB|Max_col|
+---+----+----+-------+
|ID1|   3|   5|   colB|
|ID2|   4|  12|   colB|
|ID3|   8|   3|   colA|
+---+----+----+-------+

如果出现平局，其中 colA 和 colB 具有相同的值，请选择第一列。

我怎样才能在 pyspark 中实现这一点

Answer 1

您可以UDF在每一行上使用逐行计算，并用于struct将多列传递给 udf。希望这可以帮助。

from pyspark.sql import functions as F
from pyspark.sql.types import IntegerType
from operator import itemgetter

data = [(("ID1", 3, 5,78)), (("ID2", 4, 12,45)), (("ID3", 70, 3,67))]
df = spark.createDataFrame(data, ["ID", "colA", "colB","colC"])
df.show()

+---+----+----+----+
| ID|colA|colB|colC|
+---+----+----+----+
|ID1|   3|   5|  78|
|ID2|   4|  12|  45|
|ID3|  70|   3|  70|
+---+----+----+----+
cols = df.columns

# to get max of values in a row
maxcol = F.udf(lambda row: max(row), IntegerType())
maxDF = df.withColumn("maxval", maxcol(F.struct([df[x] for x in df.columns[1:]])))
maxDF.show()

+---+----+----+----+-------+
|ID |colA|colB|colC|Max_col|
+---+----+----+----+-------+
|ID1|3   |5   |78  |78     |
|ID2|4   |12  |45  |45     |
|ID3|70  |3   |67  |70     |
+---+----+----+----+-------+

# to get max of value & corresponding column name

schema=StructType([StructField('maxval',IntegerType()),StructField('maxval_colname',StringType())])

maxcol = F.udf(lambda row: max(row,key=itemgetter(0)), schema)
maxDF = df.withColumn('maxfield', maxcol(F.struct([F.struct(df[x],F.lit(x)) for x in df.columns[1:]]))).\
select(df.columns+['maxfield.maxval','maxfield.maxval_colname'])

+---+----+----+----+------+--------------+
| ID|colA|colB|colC|maxval|maxval_colname|
+---+----+----+----+------+--------------+
|ID1| 3  | 5  | 78 | 78   | colC         |
|ID2| 4  | 12 | 45 | 45   | colC         |
|ID3| 70 | 3  | 67 | 68   | colA         |
+---+----+----+----+------+--------------+

Answer 2

有多种选择可以实现这一目标。我是一个提供例子，可以提供休息的提示-

from pyspark.sql import functions as F
from pyspark.sql.window import Window as W
from pyspark.sql import types as T

data = [(("ID1", 3, 5)), (("ID2", 4, 12)), (("ID3", 8, 3))]
df = spark.createDataFrame(data, ["ID", "colA", "colB"])
df.show()

+---+----+----+
| ID|colA|colB|
+---+----+----+
|ID1|   3|   5|
|ID2|   4|  12|
|ID3|   8|   3|
+---+----+----+

#Below F.array creates an array of column name and value pair like [['colA', 3], ['colB', 5]] then F.explode break this array into rows like different column and value pair should be in different rows

df = df.withColumn(
    "max_val",
    F.explode(
        F.array([
            F.array([F.lit(cl), F.col(cl)]) for cl in df.columns[1:]
        ])
    )
)
df.show()
+---+----+----+----------+
| ID|colA|colB|   max_val|
+---+----+----+----------+
|ID1|   3|   5| [colA, 3]|
|ID1|   3|   5| [colB, 5]|
|ID2|   4|  12| [colA, 4]|
|ID2|   4|  12|[colB, 12]|
|ID3|   8|   3| [colA, 8]|
|ID3|   8|   3| [colB, 3]|
+---+----+----+----------+

#Then select columns so that column name and value should be in different columns
df = df.select(
    "ID", 
    "colA", 
    "colB", 
    F.col("max_val").getItem(0).alias("col_name"),
    F.col("max_val").getItem(1).cast(T.IntegerType()).alias("col_value"),
)
df.show()
+---+----+----+--------+---------+
| ID|colA|colB|col_name|col_value|
+---+----+----+--------+---------+
|ID1|   3|   5|    colA|        3|
|ID1|   3|   5|    colB|        5|
|ID2|   4|  12|    colA|        4|
|ID2|   4|  12|    colB|       12|
|ID3|   8|   3|    colA|        8|
|ID3|   8|   3|    colB|        3|
+---+----+----+--------+---------+

# Rank column values based on ID in desc order
df = df.withColumn(
    "rank",
    F.rank().over(W.partitionBy("ID").orderBy(F.col("col_value").desc()))
)
df.show()
+---+----+----+--------+---------+----+
| ID|colA|colB|col_name|col_value|rank|
+---+----+----+--------+---------+----+
|ID2|   4|  12|    colB|       12|   1|
|ID2|   4|  12|    colA|        4|   2|
|ID3|   8|   3|    colA|        8|   1|
|ID3|   8|   3|    colB|        3|   2|
|ID1|   3|   5|    colB|        5|   1|
|ID1|   3|   5|    colA|        3|   2|
+---+----+----+--------+---------+----+

#Finally Filter rank = 1 as max value have rank 1 because we ranked desc value
df.where("rank=1").show()
+---+----+----+--------+---------+----+
| ID|colA|colB|col_name|col_value|rank|
+---+----+----+--------+---------+----+
|ID2|   4|  12|    colB|       12|   1|
|ID3|   8|   3|    colA|        8|   1|
|ID1|   3|   5|    colB|        5|   1|
+---+----+----+--------+---------+----+

其他选项是 -

• 在基本 df 上使用 UDF 并返回具有最大值的列名称
• 在同一示例中，在创建列名称和值列而不是排名后，使用 group by IDtake max col_value。然后与之前的df连接。