xgboost in R: how does xgb.cv pass the optimal parameters into xgb.train

Question

I\'ve been exploring the xgboost package in R and went through several demos as well as tutorials but this still confuses me: after using xgb.cv to

Answer 1

Looks like you misunderstood xgb.cv, it is not a parameter searching function. It does k-folds cross validation, nothing more.

In your code, it does not change the value of param.

To find best parameters in R's XGBoost, there are some methods. These are 2 methods,

(1) Use mlr package, http://mlr-org.github.io/mlr-tutorial/release/html/

There is a XGBoost + mlr example code in the Kaggle's Prudential challenge,

But that code is for regression, not classification. As far as I know, there is no mlogloss metric yet in mlr package, so you must code the mlogloss measurement from scratch by yourself. CMIIW.

(2) Second method, by manually setting the parameters then repeat, example,

param <- list(objective = "multi:softprob",
      eval_metric = "mlogloss",
      num_class = 12,
      max_depth = 8,
      eta = 0.05,
      gamma = 0.01, 
      subsample = 0.9,
      colsample_bytree = 0.8, 
      min_child_weight = 4,
      max_delta_step = 1
      )
cv.nround = 1000
cv.nfold = 5
mdcv <- xgb.cv(data=dtrain, params = param, nthread=6, 
                nfold=cv.nfold, nrounds=cv.nround,
                verbose = T)

Then, you find the best (minimum) mlogloss,

min_logloss = min(mdcv[, test.mlogloss.mean])
min_logloss_index = which.min(mdcv[, test.mlogloss.mean])

min_logloss is the minimum value of mlogloss, while min_logloss_index is the index (round).

You must repeat the process above several times, each time change the parameters manually (mlr does the repeat for you). Until finally you get best global minimum min_logloss.

Note: You can do it in a loop of 100 or 200 iterations, in which for each iteration you set the parameters value randomly. This way, you must save the best [parameters_list, min_logloss, min_logloss_index] in variables or in a file.

Note: better to set random seed by set.seed() for reproducible result. Different random seed yields different result. So, you must save [parameters_list, min_logloss, min_logloss_index, seednumber] in the variables or file.

Say that finally you get 3 results in 3 iterations/repeats:

min_logloss = 2.1457, min_logloss_index = 840
min_logloss = 2.2293, min_logloss_index = 920
min_logloss = 1.9745, min_logloss_index = 780

Then you must use the third parameters (it has global minimum min_logloss of 1.9745). Your best index (nrounds) is 780.

Once you get best parameters, use it in the training,

# best_param is global best param with minimum min_logloss
# best_min_logloss_index is the global minimum logloss index
nround = 780
md <- xgb.train(data=dtrain, params=best_param, nrounds=nround, nthread=6)

I don't think you need watchlist in the training, because you have done the cross validation. But if you still want to use watchlist, it is just okay.

Even better you can use early stopping in xgb.cv.

mdcv <- xgb.cv(data=dtrain, params=param, nthread=6, 
                nfold=cv.nfold, nrounds=cv.nround,
                verbose = T, early.stop.round=8, maximize=FALSE)

With this code, when mlogloss value is not decreasing in 8 steps, the xgb.cv will stop. You can save time. You must set maximize to FALSE, because you expect minimum mlogloss.

Here is an example code, with 100 iterations loop, and random chosen parameters.

best_param = list()
best_seednumber = 1234
best_logloss = Inf
best_logloss_index = 0

for (iter in 1:100) {
    param <- list(objective = "multi:softprob",
          eval_metric = "mlogloss",
          num_class = 12,
          max_depth = sample(6:10, 1),
          eta = runif(1, .01, .3),
          gamma = runif(1, 0.0, 0.2), 
          subsample = runif(1, .6, .9),
          colsample_bytree = runif(1, .5, .8), 
          min_child_weight = sample(1:40, 1),
          max_delta_step = sample(1:10, 1)
          )
    cv.nround = 1000
    cv.nfold = 5
    seed.number = sample.int(10000, 1)[[1]]
    set.seed(seed.number)
    mdcv <- xgb.cv(data=dtrain, params = param, nthread=6, 
                    nfold=cv.nfold, nrounds=cv.nround,
                    verbose = T, early.stop.round=8, maximize=FALSE)

    min_logloss = min(mdcv[, test.mlogloss.mean])
    min_logloss_index = which.min(mdcv[, test.mlogloss.mean])

    if (min_logloss < best_logloss) {
        best_logloss = min_logloss
        best_logloss_index = min_logloss_index
        best_seednumber = seed.number
        best_param = param
    }
}

nround = best_logloss_index
set.seed(best_seednumber)
md <- xgb.train(data=dtrain, params=best_param, nrounds=nround, nthread=6)

With this code, you run cross validation 100 times, each time with random parameters. Then you get best parameter set, that is in the iteration with minimum min_logloss.

Increase the value of early.stop.round in case you find out that it's too small (too early stopping). You need also to change the random parameter values' limit based on your data characteristics.

And, for 100 or 200 iterations, I think you want to change verbose to FALSE.

Side note: That is example of random method, you can adjust it e.g. by Bayesian optimization for better method. If you have Python version of XGBoost, there is a good hyperparameter script for XGBoost, https://github.com/mpearmain/BayesBoost to search for best parameters set using Bayesian optimization.

Edit: I want to add 3rd manual method, posted by "Davut Polat" a Kaggle master, in the Kaggle forum.

Edit: If you know Python and sklearn, you can also use GridSearchCV along with xgboost.XGBClassifier or xgboost.XGBRegressor

Answer 2

This is a good question and great reply from silo with lots of details! I found it very helpful for someone new to xgboost like me. Thank you. The method to randomize and compared to boundary is very inspiring. Good to use and good to know. Now in 2018 some slight revise are needed, for example, early.stop.round should be early_stopping_rounds. The output mdcv is organized slightly differently:

  min_rmse_index  <-  mdcv$best_iteration
  min_rmse <-  mdcv$evaluation_log[min_rmse_index]$test_rmse_mean

And depends on the application (linear, logistic,etc...), the objective, eval_metric and parameters shall be adjusted accordingly.

For the convenience of anyone who is running a regression, here is the slightly adjusted version of code (most are the same as above).

library(xgboost)
# Matrix for xgb: dtrain and dtest, "label" is the dependent variable
dtrain <- xgb.DMatrix(X_train, label = Y_train)
dtest <- xgb.DMatrix(X_test, label = Y_test)

best_param <- list()
best_seednumber <- 1234
best_rmse <- Inf
best_rmse_index <- 0

set.seed(123)
for (iter in 1:100) {
  param <- list(objective = "reg:linear",
                eval_metric = "rmse",
                max_depth = sample(6:10, 1),
                eta = runif(1, .01, .3), # Learning rate, default: 0.3
                subsample = runif(1, .6, .9),
                colsample_bytree = runif(1, .5, .8), 
                min_child_weight = sample(1:40, 1),
                max_delta_step = sample(1:10, 1)
  )
  cv.nround <-  1000
  cv.nfold <-  5 # 5-fold cross-validation
  seed.number  <-  sample.int(10000, 1) # set seed for the cv
  set.seed(seed.number)
  mdcv <- xgb.cv(data = dtrain, params = param,  
                 nfold = cv.nfold, nrounds = cv.nround,
                 verbose = F, early_stopping_rounds = 8, maximize = FALSE)

  min_rmse_index  <-  mdcv$best_iteration
  min_rmse <-  mdcv$evaluation_log[min_rmse_index]$test_rmse_mean

  if (min_rmse < best_rmse) {
    best_rmse <- min_rmse
    best_rmse_index <- min_rmse_index
    best_seednumber <- seed.number
    best_param <- param
  }
}

# The best index (min_rmse_index) is the best "nround" in the model
nround = best_rmse_index
set.seed(best_seednumber)
xg_mod <- xgboost(data = dtest, params = best_param, nround = nround, verbose = F)

# Check error in testing data
yhat_xg <- predict(xg_mod, dtest)
(MSE_xgb <- mean((yhat_xg - Y_test)^2))

Answer 3

I found silo's answer is very helpful. In addition to his approach of random research, you may want to use Bayesian optimization to facilitate the process of hyperparameter search, e.g. rBayesianOptimization library. The following is my code with rbayesianoptimization library.

cv_folds <- KFold(dataFTR$isPreIctalTrain, nfolds = 5, stratified = FALSE, seed = seedNum)
xgb_cv_bayes <- function(nround,max.depth, min_child_weight, subsample,eta,gamma,colsample_bytree,max_delta_step) {
param<-list(booster = "gbtree",
            max_depth = max.depth,
            min_child_weight = min_child_weight,
            eta=eta,gamma=gamma,
            subsample = subsample, colsample_bytree = colsample_bytree,
            max_delta_step=max_delta_step,
            lambda = 1, alpha = 0,
            objective = "binary:logistic",
            eval_metric = "auc")
cv <- xgb.cv(params = param, data = dtrain, folds = cv_folds,nrounds = 1000,early_stopping_rounds = 10, maximize = TRUE, verbose = verbose)

list(Score = cv$evaluation_log$test_auc_mean[cv$best_iteration],
     Pred=cv$best_iteration)
# we don't need cross-validation prediction and we need the number of rounds.
# a workaround is to pass the number of rounds(best_iteration) to the Pred, which is a default parameter in the rbayesianoptimization library.
}
OPT_Res <- BayesianOptimization(xgb_cv_bayes,
                              bounds = list(max.depth =c(3L, 10L),min_child_weight = c(1L, 40L),
                                            subsample = c(0.6, 0.9),
                                            eta=c(0.01,0.3),gamma = c(0.0, 0.2),
                                            colsample_bytree=c(0.5,0.8),max_delta_step=c(1L,10L)),
                              init_grid_dt = NULL, init_points = 10, n_iter = 10,
                              acq = "ucb", kappa = 2.576, eps = 0.0,
                              verbose = verbose)
best_param <- list(
booster = "gbtree",
eval.metric = "auc",
objective = "binary:logistic",
max_depth = OPT_Res$Best_Par["max.depth"],
eta = OPT_Res$Best_Par["eta"],
gamma = OPT_Res$Best_Par["gamma"],
subsample = OPT_Res$Best_Par["subsample"],
colsample_bytree = OPT_Res$Best_Par["colsample_bytree"],
min_child_weight = OPT_Res$Best_Par["min_child_weight"],
max_delta_step = OPT_Res$Best_Par["max_delta_step"])
# number of rounds should be tuned using CV
#https://www.hackerearth.com/practice/machine-learning/machine-learning-algorithms/beginners-tutorial-on-xgboost-parameter-tuning-r/tutorial/
# However, nrounds can not be directly derivied from the bayesianoptimization function
# Here, OPT_Res$Pred, which was supposed to be used for cross-validation, is used to record the number of rounds
nrounds=OPT_Res$Pred[[which.max(OPT_Res$History$Value)]]
xgb_model <- xgb.train (params = best_param, data = dtrain, nrounds = nrounds)