将IndexToString应用于Spark中的特征向量 [英] Applying IndexToString to features vector in Spark

问题描述

上下文:我有一个数据框，其中所有分类值都已使用StringIndexer进行了索引.

Context: I have a data frame where all categorical values have been indexed using StringIndexer.

val categoricalColumns = df.schema.collect { case StructField(name, StringType, nullable, meta) => name }    

val categoryIndexers = categoricalColumns.map {
  col => new StringIndexer().setInputCol(col).setOutputCol(s"${col}Indexed") 
}

然后我使用VectorAssembler对所有特征列(包括索引的分类特征列)进行向量化.

Then I used VectorAssembler to vectorize all feature columns (including the indexed categorical ones).

val assembler = new VectorAssembler()
    .setInputCols(dfIndexed.columns.diff(List("label") ++ categoricalColumns))
    .setOutputCol("features")

应用分类器和其他一些步骤后，我最终得到一个具有标签，功能和预测的数据框.我想将我的功能向量扩展到单独的列，以便将索引值转换回其原始String形式.

After applying the classifier and a few additional steps I end up with a data frame that has label, features, and prediction. I would like expand my features vector to separate columns in order to convert the indexed values back to their original String form.

val categoryConverters = categoricalColumns.zip(categoryIndexers).map {
colAndIndexer => new IndexToString().setInputCol(s"${colAndIndexer._1}Indexed").setOutputCol(colAndIndexer._1).setLabels(colAndIndexer._2.fit(df).labels)
}

问题:是否有简单方法来执行此操作，还是将预测列附加到测试数据帧的最佳方法?

Question: Is there a simple way of doing this, or is the best approach to somehow attach the prediction column to the test data frame?

我尝试过的事情:

val featureSlicers = categoricalColumns.map {
  col => new VectorSlicer().setInputCol("features").setOutputCol(s"${col}Indexed").setNames(Array(s"${col}Indexed"))
}

应用此选项后，我会得到想要的列，但它们都是Vector形式的(按原意这样做)，而不是Double.

Applying this gives me the columns that I want, but they are in Vector form (as it is meant to do) and not type Double.

所需的输出是原始数据帧(即分类特征为String而不是索引)，并带有附加列指示预测的标签(在我的情况下为0或1).

The desired output is the original data frame (i.e. categorical features as String not index) with an additional column indicating the predicted label (which in my case is 0 or 1).

例如，说我的分类器的输出看起来像这样:

For example, say the output of my classifier looked something like this:

+-----+---------+----------+
|label| features|prediction|
+-----+---------+----------+
|  1.0|[0.0,3.0]|       1.0|
+-----+---------+----------+

通过将VectorSlicer应用于每个功能，我会得到:

By applying VectorSlicer on each feature I would get:

+-----+---------+----------+-------------+-------------+
|label| features|prediction|statusIndexed|artistIndexed|
+-----+---------+----------+-------------+-------------+
|  1.0|[0.0,3.0]|       1.0|        [0.0]|        [3.0]|
+-----+---------+----------+-------------+-------------+

哪个很棒，但我需要:

+-----+---------+----------+-------------+-------------+
|label| features|prediction|statusIndexed|artistIndexed|
+-----+---------+----------+-------------+-------------+
|  1.0|[0.0,3.0]|       1.0|         0.0 |         3.0 |
+-----+---------+----------+-------------+-------------+

然后可以使用IndexToString并将其转换为:

To then be able to use IndexToString and convert it to:

+-----+---------+----------+-------------+-------------+
|label| features|prediction|    status   |    artist   |
+-----+---------+----------+-------------+-------------+
|  1.0|[0.0,3.0]|       1.0|        good |  Pink Floyd |
+-----+---------+----------+-------------+-------------+

甚至:

+-----+----------+-------------+-------------+
|label|prediction|    status   |    artist   |
+-----+----------+-------------+-------------+
|  1.0|       1.0|        good |  Pink Floyd |
+-----+----------+-------------+-------------+

推荐答案

嗯，这不是一个非常有用的操作，但是应该可以使用列元数据和简单的UDF提取所需的信息.我假设您的数据已创建为类似于以下内容的管道:

Well, it is not a very useful operation but it should be possible to extract required information using column metadata and as simple UDF. I assume your data has been created a pipeline similar to this one:

import org.apache.spark.ml.feature.{VectorSlicer, VectorAssembler, StringIndexer}
import org.apache.spark.ml.Pipeline

val df = sc.parallelize(Seq(
  (1L, "a", "foo", 1.0), (2L, "b", "bar", 2.0), (3L, "a", "bar", 3.0)
)).toDF("id", "x1", "x2", "x3")

val featureCols = Array("x1", "x2", "x3")
val featureColsIdx = featureCols.map(c => s"${c}_i")

val indexers = featureCols.map(
  c => new StringIndexer().setInputCol(c).setOutputCol(s"${c}_i")
)

val assembler = new VectorAssembler()
  .setInputCols(featureColsIdx)
  .setOutputCol("features")

val slicer = new VectorSlicer()
  .setInputCol("features")
  .setOutputCol("string_features")
  .setNames(featureColsIdx.init)


val transformed = new Pipeline()
  .setStages(indexers :+ assembler :+ slicer)
  .fit(df)
  .transform(df)

首先，我们可以从功能中提取所需的元数据:

First we can extract desired metadata from the features:

val meta = transformed.select($"string_features")
  .schema.fields.head.metadata
  .getMetadata("ml_attr") 
  .getMetadata("attrs")
  .getMetadataArray("nominal")

并将其转换为更易于使用的

and convert it to something easier to use

case class NominalMetadataWrapper(idx: Long, name: String, vals: Array[String])

// In general it could a good idea to make it a broadcast variable
val lookup = meta.map(m => NominalMetadataWrapper(
  m.getLong("idx"), m.getString("name"), m.getStringArray("vals")
))

最后是一个小的UDF:

Finally a small UDF:

import scala.util.Try

val transFeatures = udf((v: Vector) => lookup.map{
  m => Try(m.vals(v(m.idx.toInt).toInt)).toOption
})

transformed.select(transFeatures($"string_features")).

这篇关于将IndexToString应用于Spark中的特征向量的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！