deepImp

Imputation of missing values with configurable neural networks, for mixed-type data (impNNet()) and compositional data with rounded zeros (impNNetCoDa()). Every network head is a generalised linear model plus a neural-network correction, so the network adds non-linear effects and interactions on top of a GLM fit. Models run on a native torch backend by default (no Python required), or optionally on keras3.

• Point imputation, or multiple imputation by predictive draws (predictive mean matching or normal draws, softmax sampling for factors) on bootstrap fits, with pooling through mice (as_mids()).
• Network architecture (deepimp_arch()) and training settings (deepimp_control(): early stopping, learning-rate schedule, warm starts, parallel replicates, device) as validated objects.
• Diagnostics: summary(), convergence and chain-trace plots, per-fit training history.
• Stored models impute new data without retraining (keep_models = TRUE, predict()).

Installation

install.packages("deepImp")                      # CRAN release
# remotes::install_github("matthias-da/deepImp") # development version

torch::install_torch()     # the default backend: native libtorch, no Python

# optional: the keras backend (Python + TensorFlow)
# install.packages("keras3"); keras3::install_keras()

impNNetCoDa() also needs the robCompositions package, and as_mids() the mice package.

Quick start

Mixed-type imputation, point mode:

library(deepImp)
data(sleep, package = "VIM")
imp <- impNNet(sleep, arch = deepimp_arch_small(), epochs = 50, seed = 1)
summary(imp)
completed <- getImputed(imp)

Multiple imputation and pooled inference:

mi <- impNNet(sleep, arch = deepimp_arch_small(dropout = 0), m = 5, draw = "pmm",
              epochs = 50, seed = 1)
fit <- with(as_mids(mi), lm(Sleep ~ log(BodyWgt) + Danger))
summary(mice::pool(fit))

Compositional data (rounded zeros below a detection limit):

x <- data.frame(a = runif(50, 5, 10), b = runif(50, 5, 10), c = runif(50, 5, 10))
x$a[1:5] <- 0
imp <- impNNetCoDa(x, dl = c(1, 1, 1), label = 0, arch = deepimp_arch_small(), seed = 1)
getImputed(imp)

See vignette("deepImp") for a full walkthrough.

Reference

Templ, M. (2021). Imputation of rounded zeros for compositional data using neural networks. In: Advances in Compositional Data Analysis (Festschrift). Springer.