gbm

gbm

Generalized boosted regression models.

Classification

library(gbm)

# Binary classification
gbm_model <- gbm(
  target ~ .,
  data = train_df,
  distribution = "bernoulli",
  n.trees = 1000,
  interaction.depth = 3,
  shrinkage = 0.01
)

# Multi-class
gbm_model <- gbm(
  target ~ .,
  data = train_df,
  distribution = "multinomial",
  n.trees = 1000
)

Regression

# Gaussian (default)
gbm_model <- gbm(
  y ~ .,
  data = train_df,
  distribution = "gaussian",
  n.trees = 1000
)

# Laplace (robust)
gbm_model <- gbm(y ~ ., data = train_df, distribution = "laplace")

# Quantile regression
gbm_model <- gbm(y ~ ., data = train_df,
  distribution = list(name = "quantile", alpha = 0.5)
)

Parameters

gbm_model <- gbm(
  target ~ .,
  data = train_df,
  distribution = "bernoulli",
  n.trees = 1000,              # Number of trees
  interaction.depth = 3,       # Max tree depth
  shrinkage = 0.01,            # Learning rate
  n.minobsinnode = 10,         # Min obs in terminal node
  bag.fraction = 0.5,          # Subsample fraction
  train.fraction = 1.0,        # Training data fraction
  cv.folds = 5,                # Cross-validation folds
  keep.data = TRUE,
  verbose = TRUE
)

Optimal Trees

# With CV
gbm_model <- gbm(target ~ ., data = df, cv.folds = 5, n.trees = 5000)

# Find optimal number of trees
best_trees <- gbm.perf(gbm_model, method = "cv")

# OOB estimate
best_trees <- gbm.perf(gbm_model, method = "OOB")

# Test set
best_trees <- gbm.perf(gbm_model, method = "test")

Predictions

# Predict (returns on link scale for classification)
pred <- predict(gbm_model, newdata = test_df, n.trees = best_trees)

# Probability for classification
pred_prob <- predict(gbm_model, newdata = test_df,
                     n.trees = best_trees, type = "response")

# Multi-class probabilities
pred_prob <- predict(gbm_model, newdata = test_df,
                     n.trees = best_trees, type = "response")
# Returns array: [obs, class, 1]