[SPARK-30210] Give more informative error for BinaryClassificationEvaluator when data with only one label is provided

• Type: Improvement
• Status: Resolved
• Priority: Minor
• Resolution: Cannot Reproduce
• Affects Version/s: 3.1.0
• Component/s: ML

Environment

Pyspark on Databricks

• Assignee: Unassigned
• Reporter: Paul Anzel
• Created: 10/Dec/19 21:52
• Updated: 04/Jan/21 05:14
• Resolved: 04/Jan/21 04:05

Description

Hi all,

When I was trying to do some machine learning work with pyspark I ran into a confusing error message:

# Model and train/test set generated...
{{ evaluator = BinaryClassificationEvaluator(labelCol=label, metricName='areaUnderROC')}}
{{ prediction = model.transform(test_data)}}
{{ auc = evaluator.evaluate(prediction)}}

org.apache.spark.SparkException: Job aborted due to stage failure: Task 37 in stage 21.0 failed 4 times, most recent failure: Lost task 37.3 in stage 21.0 (TID 2811, 10.139.65.48, executor 16): java.lang.ArrayIndexOutOfBoundsException

After some investigation, I found that the issue was that the data I was trying to predict on only had one label represented, rather than both positive and negative labels. Easy enough to fix, but I would like to ask if we could replace this error with one that explicitly points out the issue. Would it be acceptable to have a check ahead of time on labels that ensures all labels are represented? Alternately, can we change the docs for BinaryClassificationEvaluator to explain what this error means?