如何在R的xgboost包中为保险索赔频率模型设置偏移变量？

Hey there! Let's break down how to implement those deductible-based offsets in your XGBoost frequency model, plus backup plans if you run into roadblocks. I'll keep this practical, since you're coming from a GLM background and need to align with your existing deductible factor logic.

First: Implementing Offsets in XGBoost #

XGBoost has a built-in parameter called base_margin that acts exactly like the offset term in GLMs. For Poisson frequency models (which I assume you're using, since claim counts are discrete), remember that the linear predictor is log(mu) = Xβ + offset. So you'll need to pass the log of your deductible factors as the base_margin—this ensures the final predicted frequency is scaled by your predefined factors, just like in your GLM setup.

Step-by-Step Code #

1. Create the offset vector (log-transformed deductible factors):

library(dplyr)

# Assume your dataset is named `insurance_data` with a `deductible` column
insurance_data <- insurance_data %>%
  mutate(
    log_deductible_factor = case_when(
      deductible == 100 ~ log(1),
      deductible == 200 ~ log(0.92),
      deductible == 300 ~ log(0.85),
      deductible == 500 ~ log(0.77),
      deductible == 700 ~ log(0.70),
      deductible == 1000 ~ log(0.64)
    )
  )

2. Train the frequency model with base_margin:

library(xgboost)

# Prepare features (exclude target, offset, and raw deductible column)
x_features <- model.matrix(~ . - claim_frequency - log_deductible_factor - deductible, 
                           data = insurance_data)
y_target <- insurance_data$claim_frequency

# Create DMatrix with base_margin set to your log-transformed factors
dtrain <- xgb.DMatrix(data = x_features, 
                      label = y_target, 
                      base_margin = insurance_data$log_deductible_factor)

# Set model parameters (Poisson objective for frequency)
freq_params <- list(
  objective = "count:poisson",
  eval_metric = "poisson-nloglik",
  learning_rate = 0.1,
  max_depth = 3,
  subsample = 0.8
)

# Train the model
xgb_freq_model <- xgb.train(
  params = freq_params,
  data = dtrain,
  nrounds = 100,
  watchlist = list(train = dtrain),
  verbose = 0
)

Why This Works #

The base_margin is added directly to XGBoost's linear predictor, so your final predicted frequency will be exp(predict(xgb_freq_model, newdata = dtest))—which equals deductible_factor * exp(Xβ), matching your GLM's output perfectly. Even with few high-deductible samples, this approach locks in your predefined factors, so you don't have to worry about the model learning noisy patterns from small subgroups.

Alternative: If base_margin Isn't Feasible #

If for some reason you can't use base_margin (e.g., compatibility issues), you can still enforce the "higher deductible = lower premium/frequency" rule with these workarounds:

1. Add Log Factor as a Feature + Monotone Constraints #

Turn your log-transformed deductible factor into a feature, then use XGBoost's monotone_constraints parameter to force the model to learn a non-negative coefficient for this feature. This guarantees that as the log factor decreases (i.e., deductible increases), the predicted frequency decreases—aligning with your business requirement.

# Add log_deductible_factor as a feature (already done in the first step)
x_features_with_factor <- model.matrix(~ . - claim_frequency - deductible, 
                                       data = insurance_data)

# Find the index of the log_deductible_factor feature
factor_col_idx <- which(colnames(x_features_with_factor) == "log_deductible_factor")

# Update params to include monotone constraint (1 = non-negative coefficient)
freq_params_with_constraint <- freq_params %>%
  modifyList(list(monotone_constraints = setNames(1, factor_col_idx)))

# Create new DMatrix and train
dtrain_with_factor <- xgb.DMatrix(data = x_features_with_factor, label = y_target)
xgb_freq_model_constrained <- xgb.train(
  params = freq_params_with_constraint,
  data = dtrain_with_factor,
  nrounds = 100,
  watchlist = list(train = dtrain_with_factor),
  verbose = 0
)

2. Weighted Training (For Extreme Small Samples) #

If high-deductible samples are extremely rare, you can assign higher weights to those observations to ensure the model doesn't ignore them. Combine this with the monotone constraint above for extra safety:

# Assign weights: e.g., 5x weight to high-deductible samples (adjust as needed)
insurance_data <- insurance_data %>%
  mutate(
    sample_weight = ifelse(deductible >= 500, 5, 1)
  )

# Add weights to DMatrix
dtrain_weighted <- xgb.DMatrix(data = x_features_with_factor, 
                               label = y_target, 
                               weight = insurance_data$sample_weight)

# Train with weighted data and monotone constraints
xgb_freq_model_weighted <- xgb.train(
  params = freq_params_with_constraint,
  data = dtrain_weighted,
  nrounds = 100,
  watchlist = list(train = dtrain_weighted),
  verbose = 0
)

Quick Note on Severity Models #

For your severity (payment amount) model, the logic is similar. Use reg:gamma or reg:squarederror (depending on your distribution) as the objective. You can apply the same base_margin approach if you have predefined severity adjustment factors for deductibles, or use monotone constraints to ensure higher deductibles lead to lower average severity (which makes sense, since smaller claims are covered by the deductible).

内容的提问来源于stack exchange，提问作者aleket