Difference between Objective and feval in xgboost

Objective

Objective in xgboost is the function which the learning algorithm will try and optimize. By definition, it must be able to create 1st (gradient) and 2nd (hessian) derivatives w.r.t. the predictions at a given training round.

A custom Objective function example:link

# user define objective function, given prediction, return gradient and second order gradient# this is loglikelihood losslogregobj <- function(preds, dtrain) {  labels <- getinfo(dtrain, "label")  preds <- 1/(1 + exp(-preds))  grad <- preds - labels  hess <- preds * (1 - preds)  return(list(grad = grad, hess = hess))}

This is the critical function to training and no xgboost model can be trained without defining one. Objective functions are directly used in splitting at each node in each tree.

feval

feval in xgboost plays no role in directly optimizing or training your model. You don't even need one to train. It doesn't impact splitting. All it does is score your model AFTER it has trained. A look at a example of a custom feval

evalerror <- function(preds, dtrain) {  labels <- getinfo(dtrain, "label")  err <- as.numeric(sum(labels != (preds > 0)))/length(labels)  return(list(metric = "error", value = err))}

Notice, it just returns a name(metric) and a score(value). Typically the feval and objective could be the same, but maybe the scoring mechanism you want is a little different, or doesn't have derivatives. For example, people use the logloss objective to train, but create an AUC feval to evaluate the model.

Furthermore you can use the feval to stop your model from training once it stops improving. And you can use multiple feval functions to score your model in different ways and observe them all.

You do not need a feval function to train a model. Only to evaluate it, and help it stop training early.

Summary:

Objective is the main workhorse.

feval is a helper to allow xgboost to do some cool things.

softmax is an objective function that is commonly used in multi-class classification. It insures that all your predictions sum to one, and are scaled using the exponential function. softmax