问题描述
我需要在Scala中内插时间序列
原始数据是2020-08-01,value1
2020-08-03,value3
我想像这样在中间日期插入数据2020-08-01,value1
2020-08-02,value2
2020-08-03,value3
其中value2是value1和value3的线性内插值
有人可以帮助我提供在Scala Spark中执行此操作的示例代码吗?由于性能原因,我宁愿避免使用UDF并使用spark.range,但我愿意接受您的最佳解决方案。
谢谢!
解决方法
0。。您可以分组并从数据框中获取最小,最大日期,并创建一个序列,将其分解以获取一系列日期。
from pyspark.sql.functions import *
from pyspark.sql import Window
w1 = Window.orderBy('date').rowsBetween(Window.unboundedPreceding,Window.currentRow)
w2 = Window.orderBy('date').rowsBetween(Window.currentRow,Window.unboundedFollowing)
df.groupBy().agg(min('date').alias('date_min'),max('date').alias('date_max')) \
.withColumn('date',sequence(to_date('date_min'),to_date('date_max'))) \
.withColumn('date',explode('date')) \
.select('date') \
.join(df,['date'],'left') \
.show(10,False)
+----------+-----+
|date |value|
+----------+-----+
|2020-08-01|0 |
|2020-08-02|null |
|2020-08-03|null |
|2020-08-04|null |
|2020-08-05|null |
|2020-08-06|10 |
+----------+-----+
1。。仅适用于您的情况,也是最简单的情况。
from pyspark.sql.functions import *
from pyspark.sql import Window
w1 = Window.orderBy('date').rowsBetween(Window.unboundedPreceding,Window.unboundedFollowing)
df.withColumn("value_m1",last('value',ignorenulls=True).over(w1)) \
.withColumn("value_p1",first('value',ignorenulls=True).over(w2)) \
.withColumn('value',coalesce(col('value'),expr('value_m1 + value_p1 / 2'))) \
.show(10,False)
+----------+-----+--------+--------+
|date |value|value_m1|value_p1|
+----------+-----+--------+--------+
|2020-08-01|0.0 |0 |0 |
|2020-08-02|5.0 |0 |10 |
|2020-08-03|10.0 |10 |10 |
+----------+-----+--------+--------+
2。。任意null
天的时间都有所改善。例如,当数据框由此给出时,
+----------+-----+
|date |value|
+----------+-----+
|2020-08-01|0 |
|2020-08-02|null |
|2020-08-03|null |
|2020-08-04|null |
|2020-08-05|null |
|2020-08-06|10 |
|2020-08-07|null |
|2020-08-08|null |
+----------+-----+
然后应按以下步骤更改代码:
from pyspark.sql.functions import *
from pyspark.sql import Window
w1 = Window.orderBy('date').rowsBetween(Window.unboundedPreceding,Window.unboundedFollowing)
w3 = Window.partitionBy('days_m1').orderBy('date')
w4 = Window.partitionBy('days_p1').orderBy(desc('date'))
df.withColumn("value_m1",ignorenulls=True).over(w2)) \
.withColumn('days_m1',count(when(col('value').isNotNull(),1)).over(w1)) \
.withColumn('days_p1',1)).over(w2)) \
.withColumn('days_m1',count(lit(1)).over(w3) - 1) \
.withColumn('days_p1',count(lit(1)).over(w4) - 1) \
.withColumn('value',expr('(days_p1 * value_m1 + days_m1 * value_p1) / (days_m1 + days_p1)'))) \
.orderBy('date') \
.show(10,False)
+----------+-----+--------+--------+-------+-------+
|date |value|value_m1|value_p1|days_m1|days_p1|
+----------+-----+--------+--------+-------+-------+
|2020-08-01|0.0 |0 |0 |0 |0 |
|2020-08-02|2.0 |0 |10 |1 |4 |
|2020-08-03|4.0 |0 |10 |2 |3 |
|2020-08-04|6.0 |0 |10 |3 |2 |
|2020-08-05|8.0 |0 |10 |4 |1 |
|2020-08-06|10.0 |10 |10 |0 |0 |
|2020-08-07|null |10 |null |1 |1 |
|2020-08-08|null |10 |null |2 |0 |
+----------+-----+--------+--------+-------+-------+