gradient-boosting-tree.R

# running time is extremely slow so we will restrict our sample keep var with 
# common mistakes : "factor quantity has new levels unknown"

df_gbt <- df %>% filter(df$construction_year!=0)
df_gbt <- df_gbt %>% filter(df_gbt$population>1)
df_gbt <- df_gbt %>% filter(df_gbt$amount_tsh>0)

# creating a new variable in the gbt data : random_100
df_gbt$random_100 <- runif(nrow(df_gbt), min=1, max=nrow(df_gbt))
# creating a new variable from the random_100 in the gbt data : subset_100
df_gbt$subset_100 <-  ntile(df_gbt$random_100, 100)

# filtering the variable subset_100 to keep only a percentage of it
df_gbt <- df_gbt %>% filter(df_gbt$subset_100 < 6)
nrow(df_gbt)

# creating a new variable in the gbt data : random
df_gbt$random <- runif(nrow(df_gbt), min=1, max=nrow(df_gbt))
# creating a new variable from the random in the gbt data : subset_100
#  which will be a part of the random  variable
df_gbt$subset <-  ntile(df_gbt$random, 10)

table(df_gbt$subset)



gbt_model_1 <- function(some_number) {
  df_train <- df_gbt %>% filter(df_gbt$subset == some_number)
  df_test <- df_gbt %>% filter(df_gbt$subset != some_number)
  
  model_gbt <- train(as.factor(status_group) ~ 
                       + gps_height + population 
                     + construction_year + longitude + latitude, 
                     data = df_train,
                     method = "xgbTree")
  
  df_test$y_pred <- predict(model_gbt, df_test)
  
  error <- sum(df_test$y_pred!=df_test$status_group)/nrow(df_test)
  
  results <- list("df_test" = df_test , "model_gbt" = model_gbt, "error" = error)
  
  return(results)
}

# generation of the data
list_results_gbt1 <- lapply(1:10, gbt_model_1) 
# printing the result and keeping each result in a vector to compute the mean
results_v = c()
for (i in 1:10){
  results_v[i] <- list_results_gbt1[[i]][3]$error
}

error_mean_gbtmodel1 = mean(results_v)
print('Average of gradient boosting tre model error rate : ')
print(error_mean_gbtmodel1)