What is ‘objective’ = ’binary:logitraw’ in XGBClassifier?

2
votes

For classification tasks with XGBoost, I know the parameter ‘objective’ = ’binary:logistic’ means specifying a binary classification task with objective function using probability. From my understanding, probability here is just calculating the positive class instances in each leaf of the decision tree.

What does then the parameter ‘objective’ = ’binary:logitraw’ do? Documentation here (https://xgboost.readthedocs.io/en/latest/parameter.html) states this means “logistic regression for binary classification, output score before logistic transformation”. What does it mean by “output score before logistic regression”, especially since probability here is calculating the positive class instances in each leaf of the decision tree, so it doesn’t make sense to me that there would be a “logistic transformation” involved?

1
pythonbinarylogistic-regressiondecision-treexgboost
cross-posted at datascience.stackexchange.com/q/81268/55122 - Ben Reiniger
Sorry about this. I was hoping to get a quicker response with higher visibility as I wasn’t getting a response from DS SE - Leockl
I am not familiar with this library or function, but logitraw probably means logits, i.e. the output before being squashed to the range [0, 1] and represent a probability. - nbro
Ok thanks @nbro. I will need to dig in deeper into this as it doesn't make sense to me how each leaf in the decision tree will have logit values. - Leockl

1 Answers

-1
votes

From documentation

binary:logitraw: logistic regression for binary classification, output score before 
logistic transformation

so the output has to go through a sigmoid to be transformed to probabilities.