#
# Library install
#
library(e1071)

#
# Calculate AUC of SVM model
#
auc_svm <- function(trained_model, variables, outcomes) {
  prediction_results <- attr(predict(trained_model,variables,probability=TRUE),'probabilities')[,'Yes']
  validation_results <- ifelse(outcomes=='Yes',1,0)
  performance(prediction(prediction_results,validation_results),'auc')@y.values[[1]]
}

#
# Perform the feature selection by 5-fold cross-validation
#
set.seed(1)
aucs <- NA
for (index in 1:folds_in_cv) {
  # Split the train_data into train_in_loop and validate_in_loop
  samp <- sample_order[(validate_count*(index-1)+1):(validate_count*index)]
  train_in_loop <- train_data[-samp,]
  validate_in_loop <- train_data[samp,]

  # Develop the model by train_in_loop
  # TODO: Remove variable one by one
  model_svm <- svm(diarrhea~age+sex+rankin_scale+tube_feeding+ppi+h2_blocker+abx,
                   probability=TRUE,
                   data=train_in_loop)

  # Calculate AUC for validate_in_loop
  auc <- auc_svm(trained_model=model_svm,
                 variables=validate_in_loop[,c('age','sex','rankin_scale','tube_feeding','ppi','h2_blocker','abx')],
                 outcomes=validate_in_loop[,'diarrhea'])
  aucs <- append(aucs, auc)
}
mean(aucs, na.rm=TRUE)

#
# Set the hyperparameter gamma and cost by 5-fold cross-validation
#
param_results <- c(0, 0, 0)
# Search gamma from 10^-5 to 10^5
for (gamma in 10^(seq(-5, 5, 0.1))) {
  # Search cost from 0.01 to 100
  for (cost in 10^(seq(-2, 2, 0.1))) {
    set.seed(1)
    aucs <- NA
    for (index in 1:folds_in_cv) {
      # Split the data into train_in_loop and validate_in_loop
      samp <- sample_order[(validate_count*(index-1)+1):(validate_count*index)]
      train_in_loop <- train_data[-samp,]
      validate_in_loop <- train_data[samp,]

      # Develop the model by train_in_loop
      model_svm <- svm(diarrhea~age+sex+rankin_scale+tube_feeding+ppi+abx, # H2RA was removed
                       gamma=gamma,
                       cost=cost,
                       probability=TRUE,
                       data=train_in_loop)

      # Calculate AUC for validate_in_loop
      auc <- auc_svm(trained_model=model_svm,
                     variables=validate_in_loop[,c('age','sex','rankin_scale','tube_feeding','ppi','abx')],
                     outcomes=validate_in_loop[,'diarrhea'])
      aucs <- append(aucs, auc)
    }
    param_results <- rbind(param_results, c(gamma, cost, mean(aucs, na.rm=TRUE)))
  }
}
head(param_results[order(param_results[,3], decreasing=TRUE),])