Uploaded image for project: 'Spark'
  1. Spark
  2. SPARK-21806

BinaryClassificationMetrics pr(): first point (0.0, 1.0) is misleading

    XMLWordPrintableJSON

Details

    • Improvement
    • Status: Resolved
    • Minor
    • Resolution: Fixed
    • 2.2.0
    • 2.3.0
    • MLlib

    Description

      I would like to reference to a discussion in scikit-learn, as this behavior is probably based on the scikit implementation.

      Summary:
      Currently, the y-axis intercept of the precision recall curve is set to (0.0, 1.0). This behavior is not ideal in certain edge cases (see example below) and can also have an impact on cross validation, when optimization metric is set to "areaUnderPR".

      Please consider blucena's post for possible alternatives.

      Edge case example:
      Consider a bad classifier, that assigns a high probability to all samples. A possible output might look like this:

      Real label Score
      1.0 1.0
      0.0 1.0
      0.0 1.0
      0.0 1.0
      0.0 1.0
      0.0 1.0
      0.0 1.0
      0.0 1.0
      0.0 1.0
      0.0 0.95
      0.0 0.95
      1.0 1.0

      This results in the following pr points (first line set by default):

      Threshold Recall Precision
      1.0 0.0 1.0
      0.95 1.0 0.2
      0.0 1.0 0,16

      The auPRC would be around 0.6. Classifiers with a more differentiated probability assignment will be falsely assumed to perform worse in regard to this auPRC.

      Attachments

        1. PRROC_example.jpeg
          51 kB
          Marc Kaminski

        Issue Links

          Activity

            People

              srowen Sean R. Owen
              MKami Marc Kaminski
              Votes:
              0 Vote for this issue
              Watchers:
              1 Start watching this issue

              Dates

                Created:
                Updated:
                Resolved: