What are different options for objective functions available in xgboost.XGBClassifier?

Question

Apart from binary:logistic (which is the default objective function), is there any other built-in objective function that can be used in xbgoost.XGBClassifier ?

Answer 1

These are the bulit-in functions available

• Objective candidate: survival:aft
• Objective candidate: binary:hinge
• Objective candidate: multi:softmax
• Objective candidate: multi:softprob
• Objective candidate: rank:pairwise
• Objective candidate: rank:ndcg
• Objective candidate: rank:map
• Objective candidate: reg:squarederror
• Objective candidate: reg:squaredlogerror
• Objective candidate: reg:logistic
• Objective candidate: binary:logistic
• Objective candidate: binary:logitraw
• Objective candidate: reg:linear
• Objective candidate: reg:pseudohubererror
• Objective candidate: count:poisson
• Objective candidate: survival:cox
• Objective candidate: reg:gamma
• Objective candidate: reg:tweedie
• Objective candidate: reg:absoluteerror

Answer 2

The default objective for XGBClassifier is ['reg:linear] however there are other parameters as well.. binary:logistic-It returns predicted probabilities for predicted class multi:softmax - Returns hard class for multiclass classification multi:softprob - It Returns probabilities for multiclass classification

Note: when using multi:softmax as objective, you need to pass num_class also as num_class is number of parameters defining number of class such as for labelliing (0,1,2), here we have 3 classes, so num_class = 3

Answer 3

That's true that binary:logistic is the default objective for XGBClassifier, but I don't see any reason why you couldn't use other objectives offered by XGBoost package. For example, you can see in sklearn.py source code that multi:softprob is used explicitly in multiclass case.

Moreover, if it's really necessary, you can provide a custom objective function (details here).