Objective function for classification

#1

I have reading the source code, specifically in the objective directory: xgboost/src/objective at master · dmlc/xgboost (github.com)

There are files and classes for regression, ranking, and multiclass classification tasks. For example the regression_obj.cu file has a RegLossObj class in it and the multiclass_obj.cu file has the SoftmaxMultiClassObj class in it (which inherits from ObjFunction). However, I cannot find a module or class structure for normal binary classification task. Would the SoftmaxMultiClassObj class just be reduced to binary classification with only two classes?

Thanks,

Karl Gardner

#2

Hi, in XGBoost there are currently two objective functions specifically designed for binary classification: one is ‘logistic’, another is ‘hinge’. The ‘hinge’ objective is wrapped up in a class called HingeObj in file hinge.cu, while ‘logistic’ objective, which works for both (probability) regression or binary classification, is contained in the file regression_obj.cu. You can find more information for available objectives in section “Learning Task Parameters” of XGBoost Parameters