Skip to contents

Multiple imputation through denoising autoencoders with dropout

Source: R/midae.R
midae.Rd

This function is used to generate multiply-imputed datasets using denoising autoencoders with Monte-Carlo dropout, early stopping and predictive mean matching (PMM).

Usage

midae(
  data,
  m = 5,
  categorical.encoding = "onehot",
  device = "cpu",
  epochs = 100,
  batch.size = 512,
  subsample = 1,
  early.stopping.epochs = 1,
  dae.params = list(),
  pmm.type = NULL,
  pmm.k = 5,
  pmm.link = "prob",
  pmm.save.vars = NULL,
  loss.na.scale = FALSE,
  verbose = TRUE,
  print.every.n = 1,
  save.model = FALSE,
  path = NULL
)

Arguments

data

A data.frame, tibble or data.table with missing values.

m

The number of imputed datasets. Default: 5.

categorical.encoding

The method for representing multi-class categorical features. Can be either "embeddings" or "onehot" (default).

device

Device to use. Either "cpu" (default) or "cuda" for GPU.

epochs

The number of training epochs (iterations). Default: 100.

batch.size

The size of samples in each batch. Default: 512.

subsample

The subsample ratio of training data. Default: 1.

early.stopping.epochs

An integer value k. The training of midae will stop if the validation performance has not improved for k epochs, only used when subsample is less than 1. Default: 1.

dae.params

A list of parameters for denoising autoencoders. See the documentation for the function dae_default() for details.

pmm.type

The type of predictive mean matching (PMM). Possible values:

  • NULL (default): Imputations without PMM;

  • 0: Imputations with PMM type 0;

  • 1: Imputations with PMM type 1;

  • 2: Imputations with PMM type 2;

  • "auto": Imputations with PMM type 2 for numeric/integer variables; imputations without PMM for categorical variables.

pmm.k

The number of donors for predictive mean matching. Default: 5

The link for predictive mean matching in binary variables:

  • "prob" (default): use probabilities;

  • "logit": use logit values.

pmm.save.vars

The names of variables whose predicted values of observed entries will be saved. Only use for PMM.

loss.na.scale

Whether to multiply the ratio of missing values in a feature to calculate the loss function. Default: FALSE.

verbose

Whether or not to print training loss information. Default: TRUE.

print.every.n

If verbose is set to TRUE, print out training loss for every n epochs. Default: 1.

save.model

Whether or not to save the imputation model. Default: FALSE.

path

The path where the final imputation model will be saved.

Examples

withNA.df <- createNA(data = iris, p = 0.2)
imputed.data <- midae(
  data = withNA.df, m = 5, epochs = 5, batch.size = 32,
  path = file.path(tempdir(), "midaemodel.pt")
)
#> [1] "cpu"
#> [1] "Running midae()."
#> Loss at epoch 1: 7.041880
#> Loss at epoch 2: 6.769073
#> Loss at epoch 3: 6.082598
#> Loss at epoch 4: 6.123292
#> Loss at epoch 5: 5.004550