Within-Fold Median Imputation (Leakage-Free)

Imputes missing values using training-fold medians only. This prevents the data leakage that occurs when medians are computed on the full dataset (including test-fold samples) before the CV loop.

**Common bug this prevents**: Computing `median(X[, j])` on the full matrix before the train/test split, which leaks test-fold information into the imputed training set.

Usage

impute_within_fold(X_train, X_test)

Arguments

X_train

Training fold feature matrix (samples x features)

X_test

Test fold feature matrix (samples x features)

Value

A list with:

X_train

Imputed training matrix (training-only medians)

X_test

Imputed test matrix (same training medians)

impute_medians

Named vector of medians used for imputation

n_imputed_train

Number of values imputed in training set

n_imputed_test

Number of values imputed in test set

Examples

if (FALSE) { # \dontrun{
for (fold in folds) {
  X_train <- feature_mat[train_idx, ]
  X_test  <- feature_mat[test_idx, ]

  # CORRECT: within-fold imputation
  imp <- impute_within_fold(X_train, X_test)
  X_train <- imp$X_train
  X_test  <- imp$X_test

  # WRONG (leaks!): imputing before the loop
  # for (j in 1:ncol(mat)) mat[is.na(mat[,j]), j] <- median(mat[,j], na.rm=TRUE)
}
} # }