How to tune hyperparameters of xgboost trees?

Question

I have a class imbalanced data & I want to tune the hyperparameters of the boosted tress using xgboost.

Questions

1. Is there an equivalent of gridsearchcv or randomsearchcv for xgboost?
2. If not what is the recommended approach to tune the parameters of xgboost?

Answer 1

Since the interface to xgboost in caret has recently changed, here is a script that provides a fully commented walkthrough of using caret to tune xgboost hyper-parameters.

For this, I will be using the training data from the Kaggle competition "Give Me Some Credit".

1. Fitting an xgboost model

In this section, we:

• fit an xgboost model with arbitrary hyperparameters
• evaluate the loss (AUC-ROC) using cross-validation (xgb.cv)
• plot the training versus testing evaluation metric

Here is some code to do this.

library(caret)
library(xgboost)
library(readr)
library(dplyr)
library(tidyr)

# load in the training data
df_train = read_csv("04-GiveMeSomeCredit/Data/cs-training.csv") %>%
  na.omit() %>%                                                                # listwise deletion 
  select(-`[EMPTY]`) %>%
  mutate(SeriousDlqin2yrs = factor(SeriousDlqin2yrs,                           # factor variable for classification
                                   labels = c("Failure", "Success")))

# xgboost fitting with arbitrary parameters
xgb_params_1 = list(
  objective = "binary:logistic",                                               # binary classification
  eta = 0.01,                                                                  # learning rate
  max.depth = 3,                                                               # max tree depth
  eval_metric = "auc"                                                          # evaluation/loss metric
)

# fit the model with the arbitrary parameters specified above
xgb_1 = xgboost(data = as.matrix(df_train %>%
                                   select(-SeriousDlqin2yrs)),
                label = df_train$SeriousDlqin2yrs,
                params = xgb_params_1,
                nrounds = 100,                                                 # max number of trees to build
                verbose = TRUE,                                         
                print.every.n = 1,
                early.stop.round = 10                                          # stop if no improvement within 10 trees
)

# cross-validate xgboost to get the accurate measure of error
xgb_cv_1 = xgb.cv(params = xgb_params_1,
                  data = as.matrix(df_train %>%
                                     select(-SeriousDlqin2yrs)),
                  label = df_train$SeriousDlqin2yrs,
                  nrounds = 100, 
                  nfold = 5,                                                   # number of folds in K-fold
                  prediction = TRUE,                                           # return the prediction using the final model 
                  showsd = TRUE,                                               # standard deviation of loss across folds
                  stratified = TRUE,                                           # sample is unbalanced; use stratified sampling
                  verbose = TRUE,
                  print.every.n = 1, 
                  early.stop.round = 10
)

# plot the AUC for the training and testing samples
xgb_cv_1$dt %>%
  select(-contains("std")) %>%
  mutate(IterationNum = 1:n()) %>%
  gather(TestOrTrain, AUC, -IterationNum) %>%
  ggplot(aes(x = IterationNum, y = AUC, group = TestOrTrain, color = TestOrTrain)) + 
  geom_line() + 
  theme_bw()

Here is what the testing versus training AUC looks like:

2. Hyperparameter search using train

For the hyperparameter search, we perform the following steps:

• create a data.frame with unique combinations of parameters that we want trained models for.
• Specify the control parameters that apply to each model's training, including the cross-validation parameters, and specify that the probabilities be computed so that the AUC can be computed
• cross-validate & train the models for each parameter combination, saving the AUC for each model.

Here is some code that shows how to do this.

# set up the cross-validated hyper-parameter search
xgb_grid_1 = expand.grid(
  nrounds = 1000,
  eta = c(0.01, 0.001, 0.0001),
  max_depth = c(2, 4, 6, 8, 10),
  gamma = 1
)

# pack the training control parameters
xgb_trcontrol_1 = trainControl(
  method = "cv",
  number = 5,
  verboseIter = TRUE,
  returnData = FALSE,
  returnResamp = "all",                                                        # save losses across all models
  classProbs = TRUE,                                                           # set to TRUE for AUC to be computed
  summaryFunction = twoClassSummary,
  allowParallel = TRUE
)

# train the model for each parameter combination in the grid, 
#   using CV to evaluate
xgb_train_1 = train(
  x = as.matrix(df_train %>%
                  select(-SeriousDlqin2yrs)),
  y = as.factor(df_train$SeriousDlqin2yrs),
  trControl = xgb_trcontrol_1,
  tuneGrid = xgb_grid_1,
  method = "xgbTree"
)

# scatter plot of the AUC against max_depth and eta
ggplot(xgb_train_1$results, aes(x = as.factor(eta), y = max_depth, size = ROC, color = ROC)) + 
  geom_point() + 
  theme_bw() + 
  scale_size_continuous(guide = "none")

Lastly, you can create a bubbleplot for the AUC over the variations of eta and max_depth:

Answer 2

Caret package have incorporated xgboost.

cv.ctrl <- trainControl(method = "repeatedcv", repeats = 1,number = 3, 
                        #summaryFunction = twoClassSummary,
                        classProbs = TRUE,
                        allowParallel=T)

    xgb.grid <- expand.grid(nrounds = 1000,
                            eta = c(0.01,0.05,0.1),
                            max_depth = c(2,4,6,8,10,14)
    )
    set.seed(45)
    xgb_tune <-train(formula,
                     data=train,
                     method="xgbTree",
                     trControl=cv.ctrl,
                     tuneGrid=xgb.grid,
                     verbose=T,
                     metric="Kappa",
                     nthread =3
    )

Sample output

eXtreme Gradient Boosting 

32218 samples
   41 predictor
    2 classes: 'N', 'Y' 

No pre-processing
Resampling: Cross-Validated (3 fold, repeated 1 times) 
Summary of sample sizes: 21479, 21479, 21478 
Resampling results

  Accuracy   Kappa      Accuracy SD   Kappa SD   
  0.9324911  0.1094426  0.0009742774  0.008972911

One drawback i see is that other parameters of xgboost like subsample etc are not supported by caret currently.

Edit

Gamma, colsample_bytree, min_child_weight and subsample etc can now (June 2017) be tuned directly using Caret. Just add them in the grid portion of the above code to make it work. Thanks usεr11852 for highliting it in the comment.

Answer 3

I know this is an old question, but I use a different method from the ones above. I use the BayesianOptimization function from the Bayesian Optimization package to find optimal parameters. To do this, you first create cross validation folds, then create a function xgb.cv.bayes that has as parameters the boosting hyper parameters you want to change. In this example I am tuning max.depth, min_child_weight, subsample, colsample_bytree, gamma. You then call xgb.cv in that function with the hyper parameters set to in the input parameters of xgb.cv.bayes. Then you call BayesianOptimization with the xgb.cv.bayes and the desired ranges of the boosting hyper parameters. init_points is the number of initial models with hyper parameters taken randomly from the specified ranges, and n_iter is the number of rounds of models after the initial points. The function outputs all boosting parameters and the test AUC.

cv_folds <- KFold(as.matrix(df.train[,target.var]), nfolds = 5, 
                  stratified = TRUE, seed = 50)
xgb.cv.bayes <- function(max.depth, min_child_weight, subsample, colsample_bytree, gamma){
  cv <- xgv.cv(params = list(booster = 'gbtree', eta = 0.05,
                             max_depth = max.depth,
                             min_child_weight = min_child_weight,
                             subsample = subsample,
                             colsample_bytree = colsample_bytree,
                             gamma = gamma,
                             lambda = 1, alpha = 0,
                             objective = 'binary:logistic',
                             eval_metric = 'auc'),
                 data = data.matrix(df.train[,-target.var]),
                 label = as.matrix(df.train[, target.var]),
                 nround = 500, folds = cv_folds, prediction = TRUE,
                 showsd = TRUE, early.stop.round = 5, maximize = TRUE,
                 verbose = 0
  )
  list(Score = cv$dt[, max(test.auc.mean)],
       Pred = cv$pred)
}

xgb.bayes.model <- BayesianOptimization(
  xgb.cv.bayes,
  bounds = list(max.depth = c(2L, 12L),
                min_child_weight = c(1L, 10L),
                subsample = c(0.5, 1),
                colsample_bytree = c(0.1, 0.4),
                gamma = c(0, 10)
  ),
  init_grid_dt = NULL,
  init_points = 10,  # number of random points to start search
  n_iter = 20, # number of iterations after initial random points are set
  acq = 'ucb', kappa = 2.576, eps = 0.0, verbose = TRUE
)

Answer 4

This is an older question but thought I would share how I tune xgboost parameters. I originally thought I would use caret for this but recently found an issue handling all of the parameters as well as missing values. I was also considering writing an iterating loop through different combinations of parameters but wanted it to run in parallel and would require too much time. Using gridSearch from the NMOF package provided the best from both worlds (all parameters as well as parallel processing). Here is example code for binary classification (works on windows and linux):

# xgboost task parameters
nrounds <- 1000
folds <- 10
obj <- 'binary:logistic'
eval <- 'logloss'

# Parameter grid to search
params <- list(
  eval_metric = eval,
  objective = obj,
  eta = c(0.1,0.01),
  max_depth = c(4,6,8,10),
  max_delta_step = c(0,1),
  subsample = 1,
  scale_pos_weight = 1
)

# Table to track performance from each worker node
res <- data.frame()

# Simple cross validated xgboost training function (returning minimum error for grid search)
xgbCV <- function (params) {
  fit <- xgb.cv(
    data = data.matrix(train), 
    label = trainLabel, 
    param =params, 
    missing = NA, 
    nfold = folds, 
    prediction = FALSE,
    early.stop.round = 50,
    maximize = FALSE,
    nrounds = nrounds
  )
  rounds <- nrow(fit)
  metric = paste('test.',eval,'.mean',sep='')
  idx <- which.min(fit[,fit[[metric]]]) 
  val <- fit[idx,][[metric]]
  res <<- rbind(res,c(idx,val,rounds))
  colnames(res) <<- c('idx','val','rounds')
  return(val)
}

# Find minimal testing error in parallel
cl <- makeCluster(round(detectCores()/2)) 
clusterExport(cl, c("xgb.cv",'train','trainLabel','nrounds','res','eval','folds'))
sol <- gridSearch(
  fun = xgbCV,
  levels = params,
  method = 'snow',
  cl = cl,
  keepNames = TRUE,
  asList = TRUE
)

# Combine all model results
comb=clusterEvalQ(cl,res)
results <- ldply(comb,data.frame)
stopCluster(cl)

# Train model given solution above
params <- c(sol$minlevels,objective = obj, eval_metric = eval)
xgbModel <- xgboost(
  data = xgb.DMatrix(data.matrix(train),missing=NaN, label = trainLabel),
  param = params,
  nrounds = results[which.min(results[,2]),1]
)

print(params)
print(results)