missknn

missknn is an R package for fast masked k-nearest neighbor imputation of missing values in tabular data.

Installation

Install the development version from GitHub with:

install.packages("remotes")
remotes::install_github("ielbadisy/missknn")

It is fast by design:

• distance computation, donor ranking, and aggregation are done in C++, with partial (not full) top-k selection and a batched per-column API that avoids per-cell R/C++ round trips
• each column’s k and estimator (distance-weighted mean vs. a local weighted-regression fit) are chosen automatically from a holdout evaluation, rather than fixed globally
• columns with no locally exploitable signal are detected and filled directly from the global mean/mode, skipping neighbor search entirely
• a donor_cap bounds the neighbor-search candidate pool so cost stays roughly linear in n instead of quadratic, which matters at large n
• data.table is used for lightweight tabular orchestration
• multiple imputation (m > 1) can run multicore by default on Unix-like systems via parallel::mclapply
• within a single imputation, the per-column holdout k/estimator tuning and the deterministic (non-stochastic) neighbor search are parallelized across columns/receivers with RcppParallel, on all platforms including Windows

It supports:

• single imputation
• multiple imputation
• numeric and mixed tabular data
• data.table-friendly workflows

About

missknn is developed and maintained by Imad El Badisy (elbadisyimad@gmail.com), released under the MIT license.

Latest Implementation

The current implementation parallelizes the per-column k/estimator tuning and the deterministic single-imputation path with RcppParallel, including on Windows. The m > 1 stochastic path still uses process-level parallelism via parallel::mclapply on Unix-like systems.

Parallel computing

missknn parallelizes at two independent levels:

• Across imputations (m > 1): each of the m completed datasets is generated in a separate process via parallel::mclapply, controlled by parallel_cores (defaults to detectCores() - 1). This only applies on Unix-like systems; it falls back to serial on Windows.
• Within a single imputation: the per-column holdout search that picks each column’s k and estimator, and the deterministic (m = 1) neighbor search/aggregation, are split across threads with RcppParallel::parallelFor - across target columns for tuning, across receiver rows for imputation. This runs on every platform, including Windows, and benefits m = 1 runs that get no benefit from parallel_cores. The stochastic sampling path used when m > 1 stays single-threaded per process, since R’s RNG isn’t thread-safe; that case still parallelizes across the m processes instead.

Thread count for the RcppParallel layer follows the usual RcppParallel::setThreadOptions() / RCPP_PARALLEL_NUM_THREADS conventions.

Benchmark

Two benchmark scripts write reports, plots, and CSV tables to inst/benchmark/output/:

# accuracy/runtime vs. missForest, missMDA, VIM::kNN at n = 200, 30 simulations
Rscript inst/benchmark/benchmark_30.R

# runtime/accuracy vs. missMDA and missForest as n scales from 1,000 to 100,000
Rscript inst/benchmark/benchmark_bigdata.R

API

missknn(data, k = 5L, m = 1L, scale = TRUE, add_indicator = FALSE, seed = NULL, weights = c("distance", "uniform"), numeric_estimator = c("regression", "mean"), ridge = 1e-4, donor_cap = 2000L, max_iter = 1L, parallel_cores = missknn_default_cores())

Fits masked KNN imputation and returns a missknn object.

Columns with no locally exploitable signal are detected during fitting and filled from the global mean/mode directly, without a neighbor search.

Returns a missknn object: a list with data (input as a data.table), completed (one completed data.table, or a list of m when m > 1), imputations (always a list of m), plus the arguments used to fit (k, m, scale, weights, etc.) and internal meta.

complete(object, action = c("all", "single"), index = 1L, ...)

S3 generic (complete.missknn) that extracts completed data from a missknn object.

• action = "all" (default): returns every completed dataset (a single data.table when m = 1, a list of m data.tables when m > 1).
• action = "single": returns one completed data.table, selected by index (ignored when m = 1, since there is only one).

print(x) / summary(object)

print.missknn reports n, p, k, m, scale, and add_indicator. summary.missknn additionally reports per-variable and total missing counts (missing_by_variable, missing_total) and has its own print.summary.missknn method.

Example

library(missknn)

x <- data.frame(
  a = c(1, 2, NA, 4),
  b = c(2, NA, 3, 5),
  g = factor(c("x", "x", "y", NA))
)

imp <- missknn(x, k = 2, m = 1)
imp
summary(imp)
complete(imp)

# multiple imputation
mi <- missknn(x, k = 2, m = 5, seed = 1)
complete(mi, action = "single", index = 2)

Execution time vs. sample size

Runtime and NRMSE from n = 1,000 up to n = 100,000 (5 numeric columns, 20% MCAR per column). missForest and VIM::kNN are only run up to n = 5,000 (per-tree cost and O(n^2) distance search respectively make larger n impractical); missknn and missRanger run across the full grid. Reproduce with:

Rscript inst/benchmark/benchmark_bigdata.R

missForest’s runtime already becomes impractical at n = 5,000 (over a minute) and would take hours at n = 100,000; VIM::kNN’s distance search is similarly infeasible past a few thousand rows. missknn’s capped-donor masked-KNN search stays close to linear in n and keeps running in well under a second even at n = 100,000. missRanger also scales to n = 100,000 but at substantially higher runtime.