Model_tuning_Script

#Linlin Cheng
#Proj 5. tuning file1

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

########################
#First round tuning:
xgb_params_1 = list(
  objective = "multi:softmax", 
  num_class = 3,
  eta = 0.1,                                                                  
  max.depth = 1, 
  eval_metric = "merror"                                                          
)

xgbtrain <- xgb.DMatrix(data.matrix(select(x_train, -lable)), label=tlabel, missing=NA)

# set up grid
xgb_grid_1 = expand.grid(
  nrounds = c(100, 500, 1000),
  eta = c(0.3, 0.1, 0.01, 0.001),
  max_depth = c(2, 4, 6, 8, 10),
  gamma = 1,
  colsample_bytree = c(0.1, 0.5),
  min_child_weight = 1 
)

# set up parameters
xgb_trcontrol_1 = trainControl(
  method = "cv",
  number = 5,
  verboseIter = TRUE,
  returnData = FALSE,
  returnResamp = "all",                                                        
  classProbs = TRUE,                                                           
  summaryFunction = multiClassSummary,
  allowParallel = TRUE
)

#tuning script:
xgb_train_1 = train(
  x = data.matrix(x_train %>% select(-lable)),
  y = make.names(as.factor(tlabel)),
  trControl = xgb_trcontrol_1,
  tuneGrid = xgb_grid_1,
  method = "xgbTree"
)

xgb_train_1$bestTune
#nrounds max_depth eta gamma colsample_bytree min_child_weight
#90    1000        10 0.1     1              0.5                1


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

#variable importance plot
plot(varImp(xgb_train_1, scale = FALSE))


#########################
#second round tuning:

#set up grid:
xgb_grid_2 = expand.grid(
  nrounds = 1000,
  eta = c(0.15, 0.1, 0.09, 0.08, 0.07),
  max_depth = c(8, 9, 10, 11),
  gamma = 1,
  colsample_bytree = c(0.4, 0.5, 0.6),
  min_child_weight = 1 
)

#set up parameters:
xgb_trcontrol_2 = trainControl(
  method = "cv",
  number = 5,
  verboseIter = TRUE,
  returnData = FALSE,
  returnResamp = "all",                                                        
  classProbs = TRUE,                                                           
  summaryFunction = multiClassSummary,
  allowParallel = TRUE
)

#model tuning:
xgb_train_2 = train(
  x = data.matrix(x_train %>% select(-lable)),
  y = make.names(as.factor(tlabel)),
  trControl = xgb_trcontrol_2,
  tuneGrid = xgb_grid_2,
  method = "xgbTree"
)

xgb_train_2$bestTune
#    nrounds max_depth  eta    gamma   colsample_bytree 
# 12 1000        11     0.07     1              0.6
#         min_child_weight
# 12                1

save(xgb_train_2, file = "xgb_tuning2.Rdata")