Skip to contents

Sliding-Window K-NN or PCA Imputation

Source: R/slide_imp.R
slide_imp.Rd

Perform sliding-window K-NN or PCA imputation on a numeric matrix whose columns are meaningfully ordered. Not intended for Illumina DNA methylation microarrays, use group_imp() instead.

Usage

slide_imp(
  obj,
  location,
  window_size,
  overlap_size = 0,
  flank = FALSE,
  min_window_n,
  subset = NULL,
  dry_run = FALSE,
  k = NULL,
  cores = 1,
  dist_pow = 0,
  ncp = NULL,
  scale = TRUE,
  coeff.ridge = 1,
  threshold = 1e-06,
  seed = NULL,
  row.w = NULL,
  nb.init = 1,
  maxiter = 1000,
  miniter = 5,
  solver = c("auto", "exact", "lobpcg"),
  lobpcg_control = NULL,
  clamp = NULL,
  method = NULL,
  .progress = TRUE,
  colmax = 0.9,
  post_imp = TRUE,
  na_check = TRUE,
  on_infeasible = c("skip", "error", "mean")
)

Arguments

obj

A numeric matrix with samples in rows and features in columns.

location

A sorted numeric vector of length ncol(obj) giving the position of each column, such as genomic coordinates.

window_size

Numeric. Window width in the same units as location.

overlap_size

Numeric. Overlap between consecutive windows in the same units as location. Must be less than window_size. Ignored when flank = TRUE.

flank

Logical. If FALSE, imputation uses sliding windows across the full matrix. If TRUE, one window of width window_size is created for each feature listed in subset; overlap_size is ignored. subset must be supplied when flank = TRUE.

min_window_n

Integer. Minimum number of columns a window must contain to be considered for imputation. For non-dry runs, the selected k or ncp value must be feasible for the usable columns in each retained window after applying colmax.

subset

Optional character or integer vector specifying columns to impute. If NULL, all eligible columns are imputed. Required when flank = TRUE.

dry_run

Logical. If TRUE, skip imputation and return a slideimp_tbl describing the windows that would be used after all filtering rules are applied. In this mode, k and ncp are not required.

k

Integer. Number of nearest neighbors to use for K-NN imputation.

cores

Integer. Number of cores to use for K-NN imputation.

dist_pow

Numeric. Power used to penalize more distant neighbors in the weighted average. dist_pow = 0 gives an unweighted average of the nearest neighbors.

ncp

Integer. Number of principal components used to predict missing entries.

scale

Logical. If TRUE, columns are scaled to unit variance.

coeff.ridge

Numeric. Ridge regularization, used only when method = "regularized". Values < 1 move toward EM PCA; values > 1 move toward mean imputation.

threshold

Numeric. Convergence threshold.

seed

Integer, numeric, or NULL. Random seed for reproducibility.

row.w

Row weights, normalized to sum to 1. NULL (equal weights), a positive numeric vector of length nrow(obj), or "n_miss" (down-weight rows with more missing values).

nb.init

Integer. Number of random initializations. The first initialization is always mean imputation.

maxiter

Integer. Maximum number of iterations.

miniter

Integer. Minimum number of iterations.

solver

Character. Eigensolver: "auto" (default), "exact", or "lobpcg". "auto" runs a short timed probe and picks "lobpcg" only when clearly faster. Consecutive EM calls warm-start LOBPCG with both the previous eigenblock and search direction. When nb.init > 1, the auto choice from the first init is reused. See Performance tips.

lobpcg_control

A list of LOBPCG eigensolver control options, usually created by lobpcg_control(). A plain named list is also accepted. Ignored when solver = "exact".

clamp

Optional numeric vector c(lower, upper) bounding PCA-imputed values (use -Inf/Inf for one-sided, NULL for none). E.g., c(0, 1) for DNAm beta values. Observed values are not clamped.

method

Character or NULL. For K-NN imputation, one of "euclidean" or "manhattan". For PCA imputation, one of "regularized" or "EM". If NULL, the corresponding backend default is used.

.progress

Logical. If TRUE, show imputation progress.

colmax

Numeric scalar between 0 and 1. Columns with a missing-data proportion greater than colmax are excluded from the main imputation method. Excluded columns are left unchanged unless post_imp = TRUE, in which case remaining missing values are replaced by column means when possible.

post_imp

Logical. If TRUE, replace missing values remaining after the main imputation method with column means when possible.

na_check

Logical. If TRUE, check whether the returned matrix still contains missing values.

on_infeasible

Character. One of "skip", "error", or "mean". Controls behavior when a window is infeasible for imputation, for example when k or ncp exceeds the number of usable columns after applying colmax.

Value

If dry_run = FALSE, a numeric matrix of the same dimensions as obj, with missing values imputed. The returned object has class slideimp_results.

If dry_run = TRUE, a data frame of class slideimp_tbl with columns start, end, and window_n, plus subset_local and, when flank = TRUE, target.

Details

The sliding-window approach divides the input matrix into smaller segments based on location values and applies imputation to each window independently. Values in overlapping regions are averaged across windows to produce the final imputed result.

Two windowing modes are supported:

  • flank = FALSE: greedily partition location into windows of width window_size with the requested overlap_size between consecutive windows.

  • flank = TRUE: create one window per feature in subset, centered on that feature using the supplied window_size.

Specify k and related arguments to use knn_imp(), or ncp and related arguments to use pca_imp().

PCA Performance tips

Speed comes from three levers: solver (through LOBPCG with warm-start), threshold, and scale. Tune these first, then accuracy parameters (ncp, coeff.ridge) on a representative subset.

Exact vs. LOBPCG with warm-start. Whether "lobpcg" beats "exact" depends on size and low-rankness: "lobpcg" is preferred for large, approximately low-rank matrices with small ncp, and "exact" for small matrices (including slide_imp() windows), where it is faster and more robust. Separately, the warm-start makes each successive solve cheap: pca_imp() warm-starts LOBPCG with the previous eigenblock and search direction, so once imputed values stabilize, later solves converge in a few iterations. The payoff therefore grows with the number of EM iterations, independent of low-rankness. solver = "auto" (default) probes both and is a safe start.

Threshold. The default 1e-6 is conservative; 1e-5 is often faster with very similar values.

Scale. For columns on a common scale (e.g., DNAm beta values in [0, 1]), scale = FALSE can be faster and more accurate.

Parallel and BLAS. In parallel via tune_imp() or group_imp() with a multithreaded BLAS, set pin_blas = TRUE to avoid thread oversubscription. On Windows, the stock BLAS can be slow. Advanced users can swap in OpenBLAS.

See Speeding up PCA imputation for the full workflow.

Examples

set.seed(1234)

# Example data with 20 samples and 100 ordered columns
beta_matrix <- sim_mat(20, 100)$input
location <- 1:100

# First perform a dry run to inspect the calculated windows
window_statistics <- slide_imp(
  beta_matrix,
  location = location,
  window_size = 50,
  overlap_size = 10,
  min_window_n = 10,
  dry_run = TRUE,
  .progress = FALSE
)
window_statistics
#> # slideimp table: 3 x 4
#>  start end window_n  subset_local
#>      1  50       50 <double [50]>
#>     41  90       50 <double [50]>
#>     81 100       20 <double [20]>

# Sliding-window K-NN imputation
imputed_knn <- slide_imp(
  beta_matrix,
  location = location,
  k = 5,
  window_size = 50,
  overlap_size = 10,
  min_window_n = 10,
  .progress = FALSE
)
imputed_knn
#> Method: slide_imp (KNN imputation)
#> Dimensions: 20 x 100
#> 
#>          feature1  feature2  feature3  feature4  feature5  feature6
#> sample1 0.3486482 0.7385414 0.4077444 0.1607935 0.3924661 0.2434143
#> sample2 0.5338935 0.4724364 0.9663621 0.4922511 0.5061132 0.3923603
#> sample3 0.7185848 0.7351035 0.6724479 0.3162537 0.7634236 1.0000000
#> sample4 0.1734418 0.0000000 0.0000000 0.0000000 0.0000000 0.2106358
#> sample5 0.4499620 0.5306182 0.5685354 0.5383513 0.4680080 0.8518388
#> sample6 0.3768380 0.5570723 0.8764909 0.5276245 0.6722794 0.5740639
#> # Showing 6 of 20 rows and 6 of 100 columns

# Sliding-window PCA imputation
imputed_pca <- slide_imp(
  beta_matrix,
  location = location,
  ncp = 2,
  window_size = 50,
  overlap_size = 10,
  min_window_n = 10,
  .progress = FALSE
)
imputed_pca
#> Method: slide_imp (PCA imputation)
#> Dimensions: 20 x 100
#> 
#>          feature1  feature2  feature3  feature4  feature5  feature6
#> sample1 0.3486482 0.7385414 0.4077444 0.1607935 0.3924661 0.2434143
#> sample2 0.5338935 0.4724364 0.9663621 0.4242777 0.5061132 0.3923603
#> sample3 0.7185848 0.7351035 0.6724479 0.3162537 0.7634236 1.0000000
#> sample4 0.1734418 0.0000000 0.0000000 0.0000000 0.0000000 0.2106358
#> sample5 0.4971181 0.5306182 0.5685354 0.5383513 0.4680080 0.8518388
#> sample6 0.3768380 0.5570723 0.8764909 0.5276245 0.6722794 0.5740639
#> # Showing 6 of 20 rows and 6 of 100 columns

# K-NN imputation with flanking windows
imputed_flank <- slide_imp(
  beta_matrix,
  location = location,
  k = 2,
  window_size = 30,
  flank = TRUE,
  subset = c(10, 30, 70),
  min_window_n = 5,
  .progress = FALSE
)
imputed_flank
#> Method: slide_imp (KNN imputation)
#> Dimensions: 20 x 100
#> 
#>          feature1  feature2  feature3  feature4  feature5  feature6
#> sample1 0.3486482 0.7385414 0.4077444 0.1607935 0.3924661 0.2434143
#> sample2 0.5338935 0.4724364 0.9663621        NA 0.5061132 0.3923603
#> sample3 0.7185848 0.7351035 0.6724479 0.3162537 0.7634236 1.0000000
#> sample4 0.1734418 0.0000000 0.0000000 0.0000000 0.0000000 0.2106358
#> sample5        NA 0.5306182 0.5685354 0.5383513 0.4680080 0.8518388
#> sample6 0.3768380 0.5570723 0.8764909 0.5276245 0.6722794 0.5740639
#> # Showing 6 of 20 rows and 6 of 100 columns