[SPARK-1213] loss function error of logistic loss - ASF JIRA

XML

Word

Printable

JSON

There might be a error in incubator-spark/mllib/src/main/scala/org/apache/spark/mllib/optimizatio/Gradient.scala

The loss function of class LogisticGradient might be wrong.

The original one is :
val loss =
if (margin > 0)

{ math.log(1 + math.exp(0 - margin)) }

else

{ math.log(1 + math.exp(margin)) - margin }

But when we use this kind of loss function, we will find that the loss is increasing when optimizing, such as LogisticRegressionWithSGD.

I think it should be something like this:

val loss =
if (label > 0)

{ math.log(1 + math.exp(margin)) }

else

{ math.log(1 + math.exp(margin)) - margin }

I tested the loss function. It works well.