When working with real-world applications, we often must contend with missing data. H2O includes a function to impute variables using the mean, median, or mode, and optionally to do so by some other grouping variables.

To examine how to impute missing data this way, we will use the small Iris dataset on flowers. In particular, we will set the petal width and length values to missing for the species "setosa" and then impute their values:

## setup iris data with some missing
d <- as.data.table(iris)
d[Species == "setosa", c("Petal.Width", "Petal.Length") := .(NA, NA)]

h2o.dmiss <- as.h2o(d, destination_frame="iris_missing")
h2o.dmeanimp <- as.h2o(d, destination_frame="iris_missing_imp")

First, we will do a simple mean imputation. This has to be done one column at a time:

## mean imputation
missing.cols <- colnames(h2o.dmiss)[apply(d, 2, anyNA)]

for (v in missing.cols) {
  h2o.dmeanimp <- h2o.impute(h2o.dmeanimp, column = v)
}

One problem with imputing the overall non-missing mean is that, if there are any systematic differences, these will be missed; also, if we could get better predictions about the missing data from any of the non-missing data, this is also missed.

Instead of a simple mean imputation, we could use a simple prediction model. The following code builds a random forest model to predict each missing column. All default values are used. If random forests take too long, a glm model could also be used:

## random forest imputation
d.imputed <- d

## prediction model
for (v in missing.cols) {
  tmp.m <- h2o.randomForest(
    x = setdiff(colnames(h2o.dmiss), v),
    y = v,
    training_frame = h2o.dmiss)
  yhat <- as.data.frame(h2o.predict(tmp.m, newdata = h2o.dmiss))
  d.imputed[[v]] <- ifelse(is.na(d.imputed[[v]]), yhat$predict, d.imputed[[v]])
}

To compare the different methods, we can create a scatter plot of petal length against petal width, with the color and shape of the points determined by the flower species. This graph has three panels. The top panel is the original data. The middle panel is the data using mean imputation. The bottom panel is the data using random forest imputation. The following code creates the graph shown in Figure 6.1:

grid.arrange(
  ggplot(iris, aes(Petal.Length, Petal.Width,
    color = Species, shape = Species)) +
    geom_point() +
    theme_classic() +
    ggtitle("Original Data"),
 ggplot(as.data.frame(h2o.dmeanimp), aes(Petal.Length, Petal.Width,
    color = Species, shape = Species)) +
    geom_point() +
    theme_classic() +
   ggtitle("Mean Imputed Data"),
 ggplot(d.imputed, aes(Petal.Length, Petal.Width,
    color = Species, shape = Species)) +
    geom_point() +
    theme_classic() +
   ggtitle("Random Forest Imputed Data"),
  ncol = 1)

Figure 6.1

In this case, the mean imputation creates aberrant values quite removed from reality. If needed, more advanced prediction models could be generated. In statistical inferences, multiple imputation is preferred over single imputation (regardless of the method) as the latter fails to account for uncertainty—that is, when imputing the missing values there is some degree of uncertainty as to exactly what those values are. However, in most use cases for deep learning, the datasets are far too large and the computational time too demanding to create multiple datasets with different imputed values, train models on each, and pool the results; thus, these simpler methods (such as mean imputation or using some other prediction model) are common.