Skip to contents

Generate a sparse dictionary with DGRDL

rs_sparse_dict_dgrdl_grid_search.Rd

[Experimental] This is the Rust implementation of dual graph regularised dictionary learning in the implementation of Pan, et al., Cell Systems, 2022. This helper function is designed to run a grid search over the data.

Usage

rs_sparse_dict_dgrdl_grid_search(
  x,
  dgrdl_params,
  seeds,
  dict_sizes,
  k_neighbours_vec,
  verbose
)

Arguments

x

Numerical matrix. Rows = samples, columns = features.

dgrdl_params

A list with the parameters for the algorithm. Missing items fall back to defaults. Expects the following items.

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

  • dict_size - Integer. Size of the dictionary. Ignored here, dict_sizes is used instead.

  • alpha - Float. Sample context regularisation weight. The higher the stronger the regularisation.

  • beta - Float. Feature context regularisation weight. The higher the stronger the regularisation.

  • max_iter - Integer. Maximum iteration for the algorithm.

  • k_neighbours - Integer. Number of k neighbours for the sample and feature Laplacian matrix. Ignored here, k_neighbours_vec is used instead.

  • admm_iter - Integer. Number of iterations for using alternating direction method of multipliers (ADMM).

  • rho - Float. ADMM step size.

seeds

Integer vector. The random seeds to include in the grid search.

dict_sizes

Integer vector. The dictionary sizes to test in the grid search.

k_neighbours_vec

Integer vector. The number of neighbours for the kNN graph generation to test in the grid search.

verbose

Boolean. Controls verbosity of the function.

Value

A list with the following elements, one entry per tested combination:

  • seed - The tested seeds.

  • dict_size - The tested dictionary sizes.

  • k_neighbours - The tested numbers of neighbours.

  • reconstruction_errs - The reconstruction errors (squared Frobenius norm) for these hyperparameters.

  • feature_laplacian_objective - The objective values of the feature Laplacian term for these hyperparameters.

  • sample_laplacian_objective - The objective values of the sample Laplacian term for these hyperparameters.