如何定义自定义聚合函数来对一列向量求和？

Question

我有一个包含两列的 DataFrame，ID 类型 Int 和 Vec 类型 Vector (org.apache.spark.mllib.linalg.Vector).

DataFrame 如下所示：

ID,Vec
1,[0,0,5]
1,[4,0,1]
1,[1,2,1]
2,[7,5,0]
2,[3,3,4]
3,[0,8,1]
3,[0,0,1]
3,[7,7,7]
....

我想做一个 groupBy($"ID") 然后通过对向量求和对每个组内的行应用聚合。

上述示例的期望输出为：

ID,SumOfVectors
1,[5,2,7]
2,[10,8,4]
3,[7,15,9]
...

可用的聚合函数将不起作用，例如df.groupBy($"ID").agg(sum($"Vec") 将导致 ClassCastException。

如何实现允许我对向量或数组求和或任何其他自定义操作的自定义聚合函数？

Answer 1

Spark >= 3.0

您可以将 Summarizer 与 sum

一起使用

import org.apache.spark.ml.stat.Summarizer

df
  .groupBy($"id")
  .agg(Summarizer.sum($"vec").alias("vec"))

Spark <= 3.0

就我个人而言，我不会为 UDAF 而烦恼。不仅冗长而且速度不快 () 相反，我会简单地使用 reduceByKey / foldByKey:

import org.apache.spark.sql.Row
import breeze.linalg.{DenseVector => BDV}
import org.apache.spark.ml.linalg.{Vector, Vectors}

def dv(values: Double*): Vector = Vectors.dense(values.toArray)

val df = spark.createDataFrame(Seq(
    (1, dv(0,0,5)), (1, dv(4,0,1)), (1, dv(1,2,1)),
    (2, dv(7,5,0)), (2, dv(3,3,4)), 
    (3, dv(0,8,1)), (3, dv(0,0,1)), (3, dv(7,7,7)))
  ).toDF("id", "vec")

val aggregated = df
  .rdd
  .map{ case Row(k: Int, v: Vector) => (k, BDV(v.toDense.values)) }
  .foldByKey(BDV.zeros[Double](3))(_ += _)
  .mapValues(v => Vectors.dense(v.toArray))
  .toDF("id", "vec")

aggregated.show

// +---+--------------+
// | id|           vec|
// +---+--------------+
// |  1| [5.0,2.0,7.0]|
// |  2|[10.0,8.0,4.0]|
// |  3|[7.0,15.0,9.0]|
// +---+--------------+

并且只是为了比较 "simple" UDAF。需要导入：

import org.apache.spark.sql.expressions.{MutableAggregationBuffer,
  UserDefinedAggregateFunction}
import org.apache.spark.ml.linalg.{Vector, Vectors, SQLDataTypes}
import org.apache.spark.sql.types.{StructType, ArrayType, DoubleType}
import org.apache.spark.sql.Row
import scala.collection.mutable.WrappedArray

Class定义：

class VectorSum (n: Int) extends UserDefinedAggregateFunction {
    def inputSchema = new StructType().add("v", SQLDataTypes.VectorType)
    def bufferSchema = new StructType().add("buff", ArrayType(DoubleType))
    def dataType = SQLDataTypes.VectorType
    def deterministic = true 

    def initialize(buffer: MutableAggregationBuffer) = {
      buffer.update(0, Array.fill(n)(0.0))
    }

    def update(buffer: MutableAggregationBuffer, input: Row) = {
      if (!input.isNullAt(0)) {
        val buff = buffer.getAs[WrappedArray[Double]](0) 
        val v = input.getAs[Vector](0).toSparse
        for (i <- v.indices) {
          buff(i) += v(i)
        }
        buffer.update(0, buff)
      }
    }

    def merge(buffer1: MutableAggregationBuffer, buffer2: Row) = {
      val buff1 = buffer1.getAs[WrappedArray[Double]](0) 
      val buff2 = buffer2.getAs[WrappedArray[Double]](0) 
      for ((x, i) <- buff2.zipWithIndex) {
        buff1(i) += x
      }
      buffer1.update(0, buff1)
    }

    def evaluate(buffer: Row) =  Vectors.dense(
      buffer.getAs[Seq[Double]](0).toArray)
}

以及用法示例：

df.groupBy($"id").agg(new VectorSum(3)($"vec") alias "vec").show

// +---+--------------+
// | id|           vec|
// +---+--------------+
// |  1| [5.0,2.0,7.0]|
// |  2|[10.0,8.0,4.0]|
// |  3|[7.0,15.0,9.0]|
// +---+--------------+

另请参阅：。

Answer 2

我建议以下（适用于 Spark 2.0.2 以上版本），它可能经过优化但非常好，您必须提前知道的一件事是创建 UDAF 实例时的矢量大小

import org.apache.spark.ml.linalg._
import org.apache.spark.mllib.linalg.WeightedSparseVector
import org.apache.spark.sql.expressions.{MutableAggregationBuffer, UserDefinedAggregateFunction}
import org.apache.spark.sql.types._

class VectorAggregate(val numFeatures: Int)
   extends UserDefinedAggregateFunction {

private type B = Map[Int, Double]

def inputSchema: StructType = StructType(StructField("vec", new VectorUDT()) :: Nil)

def bufferSchema: StructType =
StructType(StructField("agg", MapType(IntegerType, DoubleType)) :: Nil)

def initialize(buffer: MutableAggregationBuffer): Unit =
buffer.update(0, Map.empty[Int, Double])

def update(buffer: MutableAggregationBuffer, input: Row): Unit = {
    val zero = buffer.getAs[B](0)
    input match {
        case Row(DenseVector(values)) => buffer.update(0, values.zipWithIndex.foldLeft(zero){case (acc,(v,i)) => acc.updated(i, v + acc.getOrElse(i,0d))})
        case Row(SparseVector(_, indices, values)) => buffer.update(0, values.zip(indices).foldLeft(zero){case (acc,(v,i)) => acc.updated(i, v + acc.getOrElse(i,0d))}) }}
def merge(buffer1: MutableAggregationBuffer, buffer2: Row): Unit = {
val zero = buffer1.getAs[B](0)
buffer1.update(0, buffer2.getAs[B](0).foldLeft(zero){case (acc,(i,v)) => acc.updated(i, v + acc.getOrElse(i,0d))})}

def deterministic: Boolean = true

def evaluate(buffer: Row): Any = {
    val Row(agg: B) = buffer
    val indices = agg.keys.toArray.sorted
    Vectors.sparse(numFeatures,indices,indices.map(agg)).compressed
}

def dataType: DataType = new VectorUDT()
}

Answer 3

使用 pyspark 3.0.0，这是我的版本，您可以使用 Summarizer 轻松完成。您的 col 类型必须是 DenseVector

from pyspark.ml.stat import Summarizer
sdf.groupBy("ID").agg(Summarizer.mean(sdf.Vec)).show()

注意：pyspark中没有avg函数，但是可以使用mean方法

如何定义自定义聚合函数来对一列向量求和？

How to define a custom aggregation function to sum a column of Vectors?

scala

aggregate-functions

apache-spark

apache-spark-sql

apache-spark-ml