High-similarity pseudolabels from multi-domain data

Generates pseudolabels based on high cosine similarity across different domains. Specialized version for multi-domain case.

Usage

high_sim_pseudolabels(
  strata,
  k = 5,
  cos_thresh = 0.97,
  min_size = 2,
  ann_trees = 50,
  verbose = FALSE
)

Arguments

strata

List produced by multidesign::hyperdesign. Each element should have $x (samples x features matrix) and $design (metadata data.frame)

k

Number of cross-domain neighbors to examine (default: 5)

cos_thresh

Minimum cosine similarity threshold (default: 0.97)

min_size

Minimum cluster size to retain (default: 2)

ann_trees

Number of trees for RcppAnnoy index (default: 50)

verbose

Whether to print progress information (default: FALSE)

Value

Factor vector of pseudolabels with length equal to total number of samples across all domains. NA indicates samples with no confident match.

Details

This function implements the approach described in the pseudolabel documentation:

1. Pool and L2-normalize all samples across domains

2. Build approximate nearest neighbor index using angular distance

3. Find high-similarity cross-domain pairs

4. Use connected components to form clusters

5. Filter clusters by minimum size

Examples

# \donttest{
if (requireNamespace("RcppAnnoy", quietly = TRUE) &&
    requireNamespace("igraph", quietly = TRUE)) {
  set.seed(1)
  strata <- list(
    list(x = matrix(rnorm(30), 10, 3), design = data.frame(row = 1:10)),
    list(x = matrix(rnorm(30), 10, 3), design = data.frame(row = 1:10))
  )
  plabs <- high_sim_pseudolabels(strata, k = 3, cos_thresh = 0.95, ann_trees = 10)
  table(plabs, useNA = "always")
}
#> plabs
#> anchor_0005        <NA> 
#>           2          18 
# }