[SPARK-4256] Make Binary Evaluation Metrics functions defined in cases where there ar...

...e 0 positive or 0 negative examples.

Author: Andrew Bullen <andrew.bullen@workday.com>

Closes #3118 from abull/master and squashes the following commits:

c2bf2b1 [Andrew Bullen] [SPARK-4256] Update Code formatting for BinaryClassificationMetricsSpec
36b0533 [Andrew Bullen] [SYMAN-4256] Extract BinaryClassificationMetricsSuite assertions into private method
4d2f79a [Andrew Bullen] [SPARK-4256] Refactor classification metrics tests - extract comparison functions in test
f411e70 [Andrew Bullen] [SPARK-4256] Define precision as 1.0 when there are no positive examples; update code formatting per pull request comments
d9a09ef [Andrew Bullen] Make Binary Evaluation Metrics functions defined in cases where there are 0 positive or 0 negative examples.

Author: Andrew Bullen

mllib/src/main/scala/org/apache/spark/mllib/evaluation/binary/BinaryClassificationMetricComputers.scala

@@ -24,26 +24,43 @@ private[evaluation] trait BinaryClassificationMetricComputer extends Serializabl
   def apply(c: BinaryConfusionMatrix): Double
 }
 
-/** Precision. */
+/** Precision. Defined as 1.0 when there are no positive examples. */
 private[evaluation] object Precision extends BinaryClassificationMetricComputer {
-  override def apply(c: BinaryConfusionMatrix): Double =
-    c.numTruePositives.toDouble / (c.numTruePositives + c.numFalsePositives)
+  override def apply(c: BinaryConfusionMatrix): Double = {
+    val totalPositives = c.numTruePositives + c.numFalsePositives
+    if (totalPositives == 0) {
+      1.0
+    } else {
+      c.numTruePositives.toDouble / totalPositives
+    }
+  }
 }
 
-/** False positive rate. */
+/** False positive rate. Defined as 0.0 when there are no negative examples. */
 private[evaluation] object FalsePositiveRate extends BinaryClassificationMetricComputer {
-  override def apply(c: BinaryConfusionMatrix): Double =
-    c.numFalsePositives.toDouble / c.numNegatives
+  override def apply(c: BinaryConfusionMatrix): Double = {
+    if (c.numNegatives == 0) {
+      0.0
+    } else {
+      c.numFalsePositives.toDouble / c.numNegatives
+    }
+  }
 }
 
-/** Recall. */
+/** Recall. Defined as 0.0 when there are no positive examples. */
 private[evaluation] object Recall extends BinaryClassificationMetricComputer {
-  override def apply(c: BinaryConfusionMatrix): Double =
-    c.numTruePositives.toDouble / c.numPositives
+  override def apply(c: BinaryConfusionMatrix): Double = {
+    if (c.numPositives == 0) {
+      0.0
+    } else {
+      c.numTruePositives.toDouble / c.numPositives
+    }
+  }
 }
 
 /**
- * F-Measure.
+ * F-Measure. Defined as 0 if both precision and recall are 0. EG in the case that all examples
+ * are false positives.
  * @param beta the beta constant in F-Measure
  * @see http://en.wikipedia.org/wiki/F1_score
  */
@@ -52,6 +69,10 @@ private[evaluation] case class FMeasure(beta: Double) extends BinaryClassificati
   override def apply(c: BinaryConfusionMatrix): Double = {
     val precision = Precision(c)
     val recall = Recall(c)
-    (1.0 + beta2) * (precision * recall) / (beta2 * precision + recall)
+    if (precision + recall == 0) {
+      0.0
+    } else {
+      (1.0 + beta2) * (precision * recall) / (beta2 * precision + recall)
+    }
   }
 }

mllib/src/test/scala/org/apache/spark/mllib/evaluation/BinaryClassificationMetricsSuite.scala

@@ -24,39 +24,104 @@ import org.apache.spark.mllib.util.TestingUtils._
 
 class BinaryClassificationMetricsSuite extends FunSuite with MLlibTestSparkContext {
 
-  def cond1(x: (Double, Double)): Boolean = x._1 ~= (x._2) absTol 1E-5
+  private def areWithinEpsilon(x: (Double, Double)): Boolean = x._1 ~= (x._2) absTol 1E-5
 
-  def cond2(x: ((Double, Double), (Double, Double))): Boolean =
+  private def pairsWithinEpsilon(x: ((Double, Double), (Double, Double))): Boolean =
     (x._1._1 ~= x._2._1 absTol 1E-5) && (x._1._2 ~= x._2._2 absTol 1E-5)
 
+  private def assertSequencesMatch(left: Seq[Double], right: Seq[Double]): Unit = {
+      assert(left.zip(right).forall(areWithinEpsilon))
+  }
+
+  private def assertTupleSequencesMatch(left: Seq[(Double, Double)],
+       right: Seq[(Double, Double)]): Unit = {
+    assert(left.zip(right).forall(pairsWithinEpsilon))
+  }
+
+  private def validateMetrics(metrics: BinaryClassificationMetrics,
+      expectedThresholds: Seq[Double],
+      expectedROCCurve: Seq[(Double, Double)],
+      expectedPRCurve: Seq[(Double, Double)],
+      expectedFMeasures1: Seq[Double],
+      expectedFmeasures2: Seq[Double],
+      expectedPrecisions: Seq[Double],
+      expectedRecalls: Seq[Double]) = {
+
+    assertSequencesMatch(metrics.thresholds().collect(), expectedThresholds)
+    assertTupleSequencesMatch(metrics.roc().collect(), expectedROCCurve)
+    assert(metrics.areaUnderROC() ~== AreaUnderCurve.of(expectedROCCurve) absTol 1E-5)
+    assertTupleSequencesMatch(metrics.pr().collect(), expectedPRCurve)
+    assert(metrics.areaUnderPR() ~== AreaUnderCurve.of(expectedPRCurve) absTol 1E-5)
+    assertTupleSequencesMatch(metrics.fMeasureByThreshold().collect(),
+      expectedThresholds.zip(expectedFMeasures1))
+    assertTupleSequencesMatch(metrics.fMeasureByThreshold(2.0).collect(),
+      expectedThresholds.zip(expectedFmeasures2))
+    assertTupleSequencesMatch(metrics.precisionByThreshold().collect(),
+      expectedThresholds.zip(expectedPrecisions))
+    assertTupleSequencesMatch(metrics.recallByThreshold().collect(),
+      expectedThresholds.zip(expectedRecalls))
+  }
+
   test("binary evaluation metrics") {
     val scoreAndLabels = sc.parallelize(
       Seq((0.1, 0.0), (0.1, 1.0), (0.4, 0.0), (0.6, 0.0), (0.6, 1.0), (0.6, 1.0), (0.8, 1.0)), 2)
     val metrics = new BinaryClassificationMetrics(scoreAndLabels)
-    val threshold = Seq(0.8, 0.6, 0.4, 0.1)
+    val thresholds = Seq(0.8, 0.6, 0.4, 0.1)
     val numTruePositives = Seq(1, 3, 3, 4)
     val numFalsePositives = Seq(0, 1, 2, 3)
     val numPositives = 4
     val numNegatives = 3
-    val precision = numTruePositives.zip(numFalsePositives).map { case (t, f) =>
+    val precisions = numTruePositives.zip(numFalsePositives).map { case (t, f) =>
       t.toDouble / (t + f)
     }
-    val recall = numTruePositives.map(t => t.toDouble / numPositives)
+    val recalls = numTruePositives.map(t => t.toDouble / numPositives)
     val fpr = numFalsePositives.map(f => f.toDouble / numNegatives)
-    val rocCurve = Seq((0.0, 0.0)) ++ fpr.zip(recall) ++ Seq((1.0, 1.0))
-    val pr = recall.zip(precision)
+    val rocCurve = Seq((0.0, 0.0)) ++ fpr.zip(recalls) ++ Seq((1.0, 1.0))
+    val pr = recalls.zip(precisions)
     val prCurve = Seq((0.0, 1.0)) ++ pr
     val f1 = pr.map { case (r, p) => 2.0 * (p * r) / (p + r)}
     val f2 = pr.map { case (r, p) => 5.0 * (p * r) / (4.0 * p + r)}
 
-    assert(metrics.thresholds().collect().zip(threshold).forall(cond1))
-    assert(metrics.roc().collect().zip(rocCurve).forall(cond2))
-    assert(metrics.areaUnderROC() ~== AreaUnderCurve.of(rocCurve) absTol 1E-5)
-    assert(metrics.pr().collect().zip(prCurve).forall(cond2))
-    assert(metrics.areaUnderPR() ~== AreaUnderCurve.of(prCurve) absTol 1E-5)
-    assert(metrics.fMeasureByThreshold().collect().zip(threshold.zip(f1)).forall(cond2))
-    assert(metrics.fMeasureByThreshold(2.0).collect().zip(threshold.zip(f2)).forall(cond2))
-    assert(metrics.precisionByThreshold().collect().zip(threshold.zip(precision)).forall(cond2))
-    assert(metrics.recallByThreshold().collect().zip(threshold.zip(recall)).forall(cond2))
+    validateMetrics(metrics, thresholds, rocCurve, prCurve, f1, f2, precisions, recalls)
+  }
+
+  test("binary evaluation metrics for RDD where all examples have positive label") {
+    val scoreAndLabels = sc.parallelize(Seq((0.5, 1.0), (0.5, 1.0)), 2)
+    val metrics = new BinaryClassificationMetrics(scoreAndLabels)
+
+    val thresholds = Seq(0.5)
+    val precisions = Seq(1.0)
+    val recalls = Seq(1.0)
+    val fpr = Seq(0.0)
+    val rocCurve = Seq((0.0, 0.0)) ++ fpr.zip(recalls) ++ Seq((1.0, 1.0))
+    val pr = recalls.zip(precisions)
+    val prCurve = Seq((0.0, 1.0)) ++ pr
+    val f1 = pr.map { case (r, p) => 2.0 * (p * r) / (p + r)}
+    val f2 = pr.map { case (r, p) => 5.0 * (p * r) / (4.0 * p + r)}
+
+    validateMetrics(metrics, thresholds, rocCurve, prCurve, f1, f2, precisions, recalls)
+  }
+
+  test("binary evaluation metrics for RDD where all examples have negative label") {
+    val scoreAndLabels = sc.parallelize(Seq((0.5, 0.0), (0.5, 0.0)), 2)
+    val metrics = new BinaryClassificationMetrics(scoreAndLabels)
+
+    val thresholds = Seq(0.5)
+    val precisions = Seq(0.0)
+    val recalls = Seq(0.0)
+    val fpr = Seq(1.0)
+    val rocCurve = Seq((0.0, 0.0)) ++ fpr.zip(recalls) ++ Seq((1.0, 1.0))
+    val pr = recalls.zip(precisions)
+    val prCurve = Seq((0.0, 1.0)) ++ pr
+    val f1 = pr.map {
+      case (0, 0) => 0.0
+      case (r, p) => 2.0 * (p * r) / (p + r)
+    }
+    val f2 = pr.map {
+      case (0, 0) => 0.0
+      case (r, p) => 5.0 * (p * r) / (4.0 * p + r)
+    }
+
+    validateMetrics(metrics, thresholds, rocCurve, prCurve, f1, f2, precisions, recalls)
   }
 }