Question

如果我的数据包含5个类别（A，B，C，D，E）和客户数据集，其中每个客户可以属于一个，多个或没有类别。如何获取这样的数据集：

id, categories
1 , [A,C]
2 , [B]
3 , []
4 , [D,E]

并将categories列转换为一个热编码向量，如此

id, categories, encoded
1 , [A,C]     , [1,0,1,0,0]
2 , [B]       , [0,1,0,0,0]
3 , []        , [0,0,0,0,0]
4 , [D,E]     , [0,0,0,1,1]

有没有人找到一种简单的方法在火花中做到这一点？

Answer 1

要获得所需的输出，您可以使用Spark的UDF（用户定义的函数）扩展Stephen Carman答案：

// Prepare training documents from a list of (id, text, label) tuples.
val data = spark.createDataFrame(Seq(
  (0L, Seq("A", "B")),
  (1L, Seq("B")),
  (2L, Seq.empty),
  (3L, Seq("D", "E"))
)).toDF("id", "categories")

// Get distinct tags array
val tags = data
  .flatMap(r ⇒ r.getAs[Seq[String]]("categories"))
  .distinct()
  .collect()
  .sortWith(_ < _)

val cvmData = new CountVectorizerModel(tags)
  .setInputCol("categories")
  .setOutputCol("sparseFeatures")
  .transform(data)

val asDense = udf((v: Vector) ⇒ v.toDense)

cvmData
  .withColumn("features", asDense($"sparseFeatures"))
  .select("id", "categories", "features")
  .show()

这将为您提供所需的输出

+---+----------+-----------------+
| id|categories|         features|
+---+----------+-----------------+
|  0|    [A, B]|[1.0,1.0,0.0,0.0]|
|  1|       [B]|[0.0,1.0,0.0,0.0]|
|  2|        []|[0.0,0.0,0.0,0.0]|
|  3|    [D, E]|[0.0,0.0,1.0,1.0]|
+---+----------+-----------------+

Answer 2

使用CountVectorizerModel

时非常容易做到这一点

val df = spark.createDataFrame(Seq(
  (1, Seq("A","C")),
  (2, Seq("B")),
  (3, Seq()),
  (4, Seq("D","E")))
).toDF("id", "category")

val cvm = new CountVectorizerModel(Array("A","B","C","D","E"))
  .setInputCol("category")
  .setOutputCol("features")

cvm.transform(df).show()

/*
+---+--------+-------------------+
| id|category|           features|
+---+--------+-------------------+
|  1|  [A, C]|(5,[0,2],[1.0,1.0])|
|  2|     [B]|      (5,[1],[1.0])|
|  3|      []|          (5,[],[])|
|  4|  [D, E]|(5,[3,4],[1.0,1.0])|
+---+--------+-------------------+
*/

这与您想要的完全不同，但功能向量会告诉您数据中存在哪些类别。例如，在第1行中，[0,2]对应于字典的第一个和第三个元素，或者＆＃34; A＆＃34;和＆＃34; C＆＃34;在那里写的。

在Spark中从类别列表创建一个热编码向量

2 个答案: