Spark DataFrame handing empty String in OneHotEncoder

I am importing a CSV file (using spark-csv) into a DataFrame which has empty String values. When applied the OneHotEncoder, the application crashes with error requirement failed: Cannot have an empty string for name.. Is there a way I can get around this?

I could reproduce the error in the example provided on Spark ml page:

val df = sqlContext.createDataFrame(Seq(
  (0, "a"),
  (1, "b"),
  (2, "c"),
  (3, ""),         //<- original example has "a" here
  (4, "a"),
  (5, "c")
)).toDF("id", "category")

val indexer = new StringIndexer()
  .setInputCol("category")
  .setOutputCol("categoryIndex")
  .fit(df)
val indexed = indexer.transform(df)

val encoder = new OneHotEncoder()
  .setInputCol("categoryIndex")
  .setOutputCol("categoryVec")
val encoded = encoder.transform(indexed)

encoded.show()

It is annoying since missing/empty values is a highly generic case.

Thanks in advance, Nikhil

Since the OneHotEncoder/OneHotEncoderEstimator does not accept empty string for name, or you'll get the following error :

java.lang.IllegalArgumentException: requirement failed: Cannot have an empty string for name. at scala.Predef$.require(Predef.scala:233) at org.apache.spark.ml.attribute.Attribute$$anonfun$5.apply(attributes.scala:33) at org.apache.spark.ml.attribute.Attribute$$anonfun$5.apply(attributes.scala:32) [...]

This is how I will do it : (There is other way to do it, rf. @Anthony 's answer)

I'll create an UDF to process the empty category :

import org.apache.spark.sql.functions._

def processMissingCategory = udf[String, String] { s => if (s == "") "NA"  else s }

Then, I'll apply the UDF on the column :

val df = sqlContext.createDataFrame(Seq(
   (0, "a"),
   (1, "b"),
   (2, "c"),
   (3, ""),         //<- original example has "a" here
   (4, "a"),
   (5, "c")
)).toDF("id", "category")
  .withColumn("category",processMissingCategory('category))

df.show
// +---+--------+
// | id|category|
// +---+--------+
// |  0|       a|
// |  1|       b|
// |  2|       c|
// |  3|      NA|
// |  4|       a|
// |  5|       c|
// +---+--------+

Now, you can go back to your transformations

val indexer = new StringIndexer().setInputCol("category").setOutputCol("categoryIndex").fit(df)
val indexed = indexer.transform(df)
indexed.show
// +---+--------+-------------+
// | id|category|categoryIndex|
// +---+--------+-------------+
// |  0|       a|          0.0|
// |  1|       b|          2.0|
// |  2|       c|          1.0|
// |  3|      NA|          3.0|
// |  4|       a|          0.0|
// |  5|       c|          1.0|
// +---+--------+-------------+

// Spark <2.3
// val encoder = new OneHotEncoder().setInputCol("categoryIndex").setOutputCol("categoryVec")
// Spark +2.3
val encoder = new OneHotEncoderEstimator().setInputCols(Array("categoryIndex")).setOutputCols(Array("category2Vec"))
val encoded = encoder.transform(indexed)

encoded.show
// +---+--------+-------------+-------------+
// | id|category|categoryIndex|  categoryVec|
// +---+--------+-------------+-------------+
// |  0|       a|          0.0|(3,[0],[1.0])|
// |  1|       b|          2.0|(3,[2],[1.0])|
// |  2|       c|          1.0|(3,[1],[1.0])|
// |  3|      NA|          3.0|    (3,[],[])|
// |  4|       a|          0.0|(3,[0],[1.0])|
// |  5|       c|          1.0|(3,[1],[1.0])|
// +---+--------+-------------+-------------+

EDIT:

@Anthony 's solution in Scala :

df.na.replace("category", Map( "" -> "NA")).show
// +---+--------+
// | id|category|
// +---+--------+
// |  0|       a|
// |  1|       b|
// |  2|       c|
// |  3|      NA|
// |  4|       a|
// |  5|       c|
// +---+--------+

I hope this helps!