[SPARK-49615] Feature transformers are case sensitive when unintented

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.4.3
• Fix Version/s: 4.0.0
• Component/s: ML, MLlib, Spark Core
• Labels: pull-request-available
• Assignee: Weichen Xu
• Reporter: Chhavi Bansal
• Created: 12/Sep/24 12:30
• Updated: 04/Nov/24 21:41
• Resolved: 04/Nov/24 21:41

Description

Hi team,

https://spark.apache.org/docs/latest/ml-features

The feature transformers are case sensitive even though the configuration 

 

spark.conf.get("spark.sql.caseSensitive") 

 

is set to false. The user of all these transformers are forced to abide by case of the column in the dataframe

 

 val data = List(Row("the movie was great", "positive", 10, "greatest of all time"),
    Row("the movie was average", "negative", 11, "just average things, average storyline"),
    Row("movie was fun", "positive", 2, "superb screen play"))
  val schema = new StructType()
    .add("comments", StringType, true)
    .add("reviews", StringType, true)
    .add("counts", IntegerType, true)
    .add("Additional_COMMENTS", StringType, true)
val df = spark.createDataFrame(data.asJava, schema)
  val si = new StringIndexer().setInputCol("additional_comments").setOutputCol("si_additional_comments")
  si.fit(df).transform(df).show() 

The above code fails with 

Exception in thread "main" org.apache.spark.SparkException: Input column additional_comments does not exist.
    at org.apache.spark.ml.feature.StringIndexerBase.$anonfun$validateAndTransformSchema$2(StringIndexer.scala:128)
    at scala.collection.TraversableLike.$anonfun$flatMap$1(TraversableLike.scala:245)
    at scala.collection.IndexedSeqOptimized.foreach(IndexedSeqOptimized.scala:36)
    at scala.collection.IndexedSeqOptimized.foreach$(IndexedSeqOptimized.scala:33)
    at scala.collection.mutable.ArrayOps$ofRef.foreach(ArrayOps.scala:198)
    at scala.collection.TraversableLike.flatMap(TraversableLike.scala:245)
    at scala.collection.TraversableLike.flatMap$(TraversableLike.scala:242)
    at scala.collection.mutable.ArrayOps$ofRef.flatMap(ArrayOps.scala:198)
    at org.apache.spark.ml.feature.StringIndexerBase.validateAndTransformSchema(StringIndexer.scala:123)
    at org.apache.spark.ml.feature.StringIndexerBase.validateAndTransformSchema$(StringIndexer.scala:115) 

Which means that the column "additional_comments" needs to be provided in the same case as in the dataframe. 

 

I think when the caseSensitive  setting is set to false we should be able to use the naming in any case.

 

Can someone please help to solve this bug for all transformers.?

Issue Links

• links to: GitHub Pull Request #48398
• links to: GitHub Pull Request #48747