如何获取xgb.DMatrix索引或添加信息以用于feval函数调参？

Great question—both approaches you’ve outlined are valid and effective ways to associate additional sample information with your custom feval function in XGBoost for R. Let’s break down how each method works and provide complete implementations:

1. Attaching Custom Metadata Directly to xgb.DMatrix #

This is the most straightforward approach, exactly as you’ve started. You can store any custom data (like your ID field) directly in the xgb.DMatrix object during creation, then retrieve it within your feval function using getinfo().

Full Implementation Example #

library(xgboost)
library(data.table)

# Sample input data
OBSERVATIONS <- data.table(
  actual = rnorm(100),  # Target values
  ID = paste0("ID_", 1:100)  # Unique sample IDs
)
scalingFact <- data.table(
  ID = paste0("ID_", 1:100),  # Matching IDs
  scale = runif(100, 0.8, 1.2)  # Extra metadata to use in evaluation
)
feature_matrix <- matrix(rnorm(100 * 10), nrow = 100)  # Model features

# Create DMatrix with custom ID metadata
dtrain <- xgb.DMatrix(
  data = feature_matrix,
  label = OBSERVATIONS$actual,
  ID = OBSERVATIONS$ID  # Attach custom field
)

# Complete custom evaluation function
custom_feval <- function(preds, dtrain) {
  # Extract label and custom metadata from DMatrix
  actual_values <- getinfo(dtrain, "label")
  sample_ids <- getinfo(dtrain, "ID")
  
  # Combine predictions, actuals, and IDs into a data table
  eval_data <- data.table(
    ID = sample_ids,
    actual = actual_values,
    preds = preds
  )
  
  # Join with your additional metadata
  eval_data <- scalingFact[eval_data, on = "ID"]
  
  # Calculate your custom metric (example: scaled MAE)
  scaled_mean_abs_error <- mean(abs(eval_data$actual - eval_data$preds) * eval_data$scale)
  
  # Return metric in the format XGBoost expects
  return(list(metric = "scaled_mae", value = scaled_mean_abs_error))
}

# Test with cross-validation
xgb.cv(
  data = dtrain,
  nrounds = 10,
  nfold = 5,
  feval = custom_feval,
  verbose = 1
)

2. Accessing Sample Indices from xgb.DMatrix #

You can also retrieve the indices of samples in the xgb.DMatrix to link back to your original dataset. By default, XGBoost assigns indices starting at 0, but you can explicitly set custom indices (like your original data’s row numbers) when creating the DMatrix for clearer alignment.

Full Implementation Example #

# Create DMatrix with explicit sample indices (using original row numbers)
dtrain <- xgb.DMatrix(
  data = feature_matrix,
  label = OBSERVATIONS$actual,
  index = 1:nrow(OBSERVATIONS)  # Use original data's row indices
)

# Custom evaluation function using indices
custom_feval_with_indices <- function(preds, dtrain) {
  # Extract indices and target values from DMatrix
  sample_indices <- getinfo(dtrain, "index")
  actual_values <- getinfo(dtrain, "label")
  
  # Fetch sample IDs and other metadata using indices from original data
  eval_data <- OBSERVATIONS[sample_indices, .(ID, actual = actual_values)]
  eval_data$preds <- preds
  
  # Join with additional metadata
  eval_data <- scalingFact[eval_data, on = "ID"]
  
  # Calculate custom metric
  scaled_mean_abs_error <- mean(abs(eval_data$actual - eval_data$preds) * eval_data$scale)
  
  return(list(metric = "scaled_mae", value = scaled_mean_abs_error))
}

# Test the function
xgb.cv(
  data = dtrain,
  nrounds = 10,
  nfold = 5,
  feval = custom_feval_with_indices,
  verbose = 1
)

Which Approach Should You Use? #

• Use the custom metadata method if you want to store the exact information you need directly in the DMatrix—this avoids relying on external data row order and makes your feval function self-contained.
• Use the index method if you prefer to keep metadata in a separate dataset and just need a way to map predictions back to their original samples.

Both methods fully meet your requirement of linking extra observation information to your custom evaluation logic.

内容的提问来源于stack exchange，提问作者P_Sta