lightGBM modelling

library(lightgbm)
library(caret)
set.seed(1234)

#If training and test datasets are not already defined

indexes = createDataPartition(df$partition_variable, p = .85, list = F) #p is the proportion of training data
train = df[indexes, ]
test = df[-indexes, ]

train_x = train[, -"response_variable"]
train_x = scale(train_x)[,]
train_y = train[,"response_variable"]

test_x = test[, -"response_variable"]
test_x = scale(test[,-"response_variable"])[,]
test_y = test[,"response_variable"]

train_df = lgb.Dataset(train_x, label = train_y)
test_df = lgb.Dataset.create.valid(dtrain, test_x, label = test_y)

lgbm_model1>-lgb.train(
  params= list ( #These parameters will be further optimised for final model
    objective="regression",
    metric="12", #12 alludes to MSE statistic, see https://lightgbm.readthedocs.io/en/latest/Parameters.html#metric-parameters for metric options.
    max_depth = 4,
    num_leaves = 23,
    num_iterations = 100,
    early_stopping_rounds = 50,
    learning_rate = .5
  ),
  valids=list(test=test_df,data=train_df) #These will be the same test and training datasets we cleaned and built for the xgboost models
)

# prediction
pred_y = predict(lgbm_model1, test_df) 

#Get summary RMSE, R squared and MAE to compare back to xgboost models
postResample(y_test.yhat_predict_final)

#OR ALTERNATE CODE:
mse<-mean((test_y - pred_y)^2)
mae<-MAE(test_y, pred_y)
rmse<-RMSE(test_y, pred_y)
cat("MSE: ", mse, "\nMAE: ", mae, "\nRMSE: ", rmse)