Skip to contents

Grid search over DGRDL parameters

dgrdl_grid_search.Rd

This function allows you to quickly iterate over different initial seeds, number of neighbours for the KNN graph and dictionary sizes to identify optimal hyperparameters for your DGRDL run.

Usage

dgrdl_grid_search(
  object,
  neighbours_vec,
  dict_size_vec,
  seed_vec,
  dgrdl_params = params_dgrdl(),
  .verbose = TRUE
)

Arguments

object

The class, see BulkCoExp(). Ideally, you should run preprocess_bulk_coexp() before applying this function.

neighbours_vec

Integer vector. The different k nearest neighbours to test.

dict_size_vec

Integer vector. The different dictionary sizes to test for.

seed_vec

Integer vector. The different initial seeds to test for the dictionary generation.

dgrdl_params

List. Output of params_dgrdl():

  • sparsity - Integer. Sparsity constraint (max non-zero coefficients per signal)

  • dict size - Integer. Will be ignored by this function and the dict_size_vec vector will be used.

  • alpha - Float. Sample context regularisation weight.

  • beta - Float. Feature effect regularisation weight.

  • max_iter - Integer. Maximum number of iterations for the main algorithm.

  • k_neighbours - Integer. Will be ignored by this function and the neighbours_vec will be used.

  • admm_iter - Integer. ADMM iterations for sparse coding.

  • rho - Float. ADMM step size.

.verbose

Boolean. Controls verbosity of the function.

Value

BulkCoExp with the grid search results added to the class.

References

Pan et al., Cell Syst, 2022