updated LogLoss (boosting) for numerical stability

apache · Nov 26, 2014 · ed5da2c · ed5da2c
1 parent 5e52bff
commit ed5da2c
Showing 1 changed file with 2 additions and 1 deletion.
diff --git a/mllib/src/main/scala/org/apache/spark/mllib/tree/loss/LogLoss.scala b/mllib/src/main/scala/org/apache/spark/mllib/tree/loss/LogLoss.scala
@@ -60,7 +60,8 @@ object LogLoss extends Loss {
   override def computeError(model: TreeEnsembleModel, data: RDD[LabeledPoint]): Double = {
     data.map { case point =>
       val prediction = model.predict(point.features)
-      2.0 * math.log(1 + math.exp(-2.0 * point.label * prediction))
+      // Use log1p since it is more stable than explicitly writing log(1 + exp()).
+      2.0 * math.log1p(math.exp(-2.0 * point.label * prediction))
     }.mean()
   }
 }