knn_imp() no longer caches pair-wise distances. The
use_cache argument has been removed.pca_imp() gained a solver argument to
select the LOBPCG eigensolver, configured via the new
lobpcg_control() helper.
pca_imp() gained a clamp argument to
bound imputed values, e.g. c(0, 1) for methylation beta
values. Also threaded through slide_imp() and
group_imp().
knn_imp() now uses {RcppThread} instead
of OpenMP, enabling parallelism on macOS.
tune_imp() now infers the subset from
na_loc, speeding up tuning for knn_imp() and
slide_imp().
prep_groups() is now an S3 generic, replacing the
previous register-on-load pattern with
{slideimp.extra}.
New article describing how to speed up PCA imputation by choosing
solver and threshold.
Fixed numerical tolerance check failure on CRAN’s ATLAS configuration.
group_imp() enforces stricter data validation. The
requested feature subset must be a subset the object’s column names
which must be a subset of the mapping data.frame. Set
allow_unmapped = TRUE to bypass errors when intersections
are incomplete.
group_imp() and tune_imp() now error
when arguments are supplied that do not apply to the chosen imputation
method, rather than silently ignoring them.
knn_imp() now uses a logical tree
argument to toggle between Ball tree (TRUE) and brute force
(FALSE). KD tree is no longer supported.
knn_imp() and pca_imp() gain more early
errors and early exits.
pca_imp() gains the same colmax and
post_imp arguments as knn_imp().
prep_groups() (formerly
group_features()) is the new name for the grouping
function. It now accepts a column name vector instead of a full
matrix.
sample_na_loc() (formerly inject_na())
is now exported. The original remains accessible via
slideimp:::inject_na() for legacy code.
sim_mat() now returns a matrix in sample-by-column
format for immediate compatibility with other package functions.
perc_NA is renamed to perc_total_na, and
dimensions are now specified via n (rows) and
p (columns).
tune_imp() gains a unified method
argument that applies to both pca_imp() and
knn_imp(), replacing pca_method and
knn_method. The rep argument is renamed to
n_reps.
tune_imp() results from v0.5.4 are no longer
reproducible because internal NA generation now uses
sample_na_loc().
The khanmiss1 dataset has been removed.
compute_metrics() now supports data frames with a
result list column containing truth and estimate columns,
similar to {yardstick}.
group_imp() and prep_groups()
automatically look up Illumina manifests using the register-on-load
pattern for {slideimp.extra}.
knn_imp() gains max_cache to control
the internal cache size (defaults to 4GB).
sim_mat() gains a rho argument to
support compound symmetry correlation structures in simulated
matrices.
sim_mat() and tune_imp() gain dedicated
print methods that provide concise summaries instead of dumping raw data
to the console.
slide_imp() gains location,
flank, and dry_run arguments for fixed-window
imputation, “flank mode” for features surrounding a subset, and
pre-computation inspection of window statistics.
tune_imp() gains granular control over NA injection
via n_cols, n_rows, num_na, and
na_col_subset. Pre-calculated locations can also be passed
to na_loc to compare methods using identical NA
patterns.
col_vars() and mean_imp_col() have been
overhauled to use the faster {RcppArmadillo} backend and
now support parallel computation with OpenMP.
Dependencies are streamlined. {tibble} and
{purrr} are removed as hard dependencies,
{cli} is added for more informative messaging, and
{carrier} is added as an explicit dependency.
Documentation is thoroughly overhauled with numerous consistency improvements and bug fixes.
{RhpcBLASctl} is added as a suggested package to
allow pinning BLAS cores and avoid thrashing during parallel
runs.
group_imp() and tune_imp() prioritize
process-level parallelization via {mirai}.
knn_imp() supports OpenMP-controlled parallelization via
the cores argument when {mirai} daemons are
not active.
knn_imp() and pca_imp() use optimized
internal Rcpp functions for better performance.
CRAN resubmission.
group_features() is added to help with creating the
group tibble needed for group_imp().
pca_imp() now allows row.w = "n_miss"
to scale row weights by the number of missing values per row.