Sparsify Embedding Matrix — sparsify

Applies sparsification to an embedding matrix by zeroing out values between specified quantiles. Includes fallback strategies if initial sparsification results in all zeros.

Usage

sparsify_embeddings(
  embedding_matrix,
  lower_quantile = 0.025,
  upper_quantile = 0.975,
  fallback_lower = 0.1,
  fallback_upper = 0.9
)

Arguments

embedding_matrix: Numeric matrix with items as columns, dimensions as rows
lower_quantile: Lower quantile threshold (default 0.025)
upper_quantile: Upper quantile threshold (default 0.975)
fallback_lower: Fallback lower quantile if first attempt fails (default 0.10)
fallback_upper: Fallback upper quantile if first attempt fails (default 0.90)

Value

Sparsified embedding matrix with same dimensions as input

Details

Sparsification process:

Zero out values between lower and upper quantiles
If result is all zeros, try fallback quantiles
If still all zeros, return original matrix

silently is always TRUE. It is only set to FALSE for developement and diagnostic purposes.