Skip to contents

Create Train/Validation Splits for LSTM Sequences

Source: R/get_lstm_sequence_splits.R
get_lstm_sequence_splits.Rd

This function creates train/validation splits from LSTM sequences using either random sampling or location-month blocking. The location-month blocking approach groups sequences by country-month combinations and randomly assigns entire blocks to train/test, which better preserves spatial-temporal structure and tests generalization capability.

Usage

get_lstm_sequence_splits(
  train_sequences,
  split_method = "random",
  train_prop = 0.6,
  seed = 99,
  verbose = TRUE
)

Arguments

train_sequences

List containing LSTM sequences with components X, y, countries, and dates (the sequence structure built internally by the legacy est_suitability() path).

split_method

Character. Either "random" (default) or "location_month_block"

train_prop

Numeric. Proportion of data for training (default 0.6)

seed

Integer. Random seed for reproducibility

verbose

Logical. Whether to print detailed information about splits (default TRUE)

Value

List containing train/validation indices and metadata:

  • train_indices: Indices for training sequences

  • val_indices: Indices for validation sequences

  • split_info: List with split statistics and metadata

Details

Random Method (existing approach):

  • Randomly samples individual sequences for train/test split

  • Good for general model development and comparison

  • May allow data leakage for sequences from same location-month

Location-Month Block Method (new approach):

  • Groups sequences by iso_code + year-month combinations

  • Randomly assigns entire location-month blocks to train or validation

  • Reduces, but does NOT eliminate, train/validation leakage: each sequence carries a multi-week input window, so sequences in adjacent location-month blocks share overlapping inputs and autocorrelated targets across the split. Validation scores are therefore optimistic relative to a strictly forward-in-time (rolling-origin) evaluation.

  • Better tests spatial-temporal generalization than a per-sequence random split

The location-month blocking approach is particularly valuable for:

  • Testing model generalization to unseen location-time combinations

  • Preventing overfitting to specific location-month patterns

  • Evaluating suitability for early warning systems

  • Assessing transferability across different epidemiological contexts

Examples

if (FALSE) { # \dontrun{
# train_sequences is a list with X, y, countries, and dates components
# (built internally by the legacy est_suitability() path).

# Random splitting (existing method)
splits_random <- get_lstm_sequence_splits(
  train_sequences,
  split_method = "random",
  train_prop = 0.6,
  seed = 99
)

# Location-month blocking (new method)
splits_blocked <- get_lstm_sequence_splits(
  train_sequences,
  split_method = "location_month_block",
  train_prop = 0.6,
  seed = 99
)

# Use the splits in LSTM training
X_train <- train_sequences$X[splits_blocked$train_indices, , ]
y_train <- train_sequences$y[splits_blocked$train_indices]
X_val <- train_sequences$X[splits_blocked$val_indices, , ]
y_val <- train_sequences$y[splits_blocked$val_indices]

# Check split information
print(splits_blocked$split_info)
} # }