Spark,DataFrame:在组上应用变换器/估计器

Question

我有一个看起来如下的DataFrame:

+-----------+-----+------------+
|     userID|group|    features|
+-----------+-----+------------+
|12462563356|    1|  [5.0,43.0]|
|12462563701|    2|   [1.0,8.0]|
|12462563701|    1|  [2.0,12.0]|
|12462564356|    1|   [1.0,1.0]|
|12462565487|    3|   [2.0,3.0]|
|12462565698|    2|   [1.0,1.0]|
|12462565698|    1|   [1.0,1.0]|
|12462566081|    2|   [1.0,2.0]|
|12462566081|    1|  [1.0,15.0]|
|12462566225|    2|   [1.0,1.0]|
|12462566225|    1|  [9.0,85.0]|
|12462566526|    2|   [1.0,1.0]|
|12462566526|    1|  [3.0,79.0]|
|12462567006|    2| [11.0,15.0]|
|12462567006|    1| [10.0,15.0]|
|12462567006|    3| [10.0,15.0]|
|12462586595|    2|  [2.0,42.0]|
|12462586595|    3|  [2.0,16.0]|
|12462589343|    3|   [1.0,1.0]|
+-----------+-----+------------+

列类型为:userID:Long,group:Int和features:vector.

这已经是分组的DataFrame,即userID将最多一次出现在特定的组中.

我的目标是features按组扩展列.

有没有办法在每个组中应用特征变换器(在我的情况下我想应用StandardScaler)而不是将其应用于完整的DataFrame.

使用ML的PS不是强制性的,因此如果解决方案基于MLlib则没有问题.

Answer 1

您可以使用与默认代码几乎相同的代码按组计算统计信息Scaler:

import org.apache.spark.mllib.stat.MultivariateOnlineSummarizer
import org.apache.spark.mllib.linalg.{Vector, Vectors}
import org.apache.spark.sql.Row

// Compute Multivariate Statistics 
val summary = data.select($"group", $"features")
    .rdd
    .map {
         case Row(group: Int, features: Vector) => (group, features) 
    }
    .aggregateByKey(new MultivariateOnlineSummarizer)(/* Create an empty new MultivariateOnlineSummarizer */
         (agg, v) => agg.add(v), /* seqOp : Add a new sample Vector to this summarizer, and update the statistical summary. */
         (agg1, agg2) => agg1.merge(agg2)) /* combOp : As MultivariateOnlineSummarizer accepts a merge action with another MultivariateOnlineSummarizer, and update the statistical summary. */
    .mapValues {
      s => (
         s.variance.toArray.map(math.sqrt(_)), /* compute the square root variance for each key */
         s.mean.toArray /* fetch the mean for each key */
      )
    }.collectAsMap

如果预期的群组数量相对较低,您可以广播这些:

val summaryBd = sc.broadcast(summary)

并转换您的数据:

val scaledRows = df.map{ case Row(userID, group: Int, features: Vector) =>
  val (stdev, mean)  =  summaryBd.value(group)
  val vs = features.toArray.clone()
  for (i <- 0 until vs.size) {
    vs(i) = if(stdev(i) == 0.0) 0.0 else (vs(i) - mean(i)) * (1 / stdev(i))
  }
  Row(userID, group, Vectors.dense(vs))
}
val scaledDf = sqlContext.createDataFrame(scaledRows, df.schema)

否则你可以简单地加入.将此列包装为具有组列作为参数的ML变换器应该不难.