Specifying One-hot Encoding on Combinations in Tidymodels - tidymodels

Question

I have a data set with two columns Value and Pattern. Value takes one of two values 0 or 1. The Pattern column is a three-character string. The three-character string or Pattern is derived from 1 of 7 characters (X, 4, 3, 2, 1, 0, T) the characters can repeat so there are (7 x 7 x 7) = 343 possible Patterns.
I would like to use tidymodels to do some classification modeling and was planning on using one-hot encoding on the Pattern column, but have the following questions:
Is there a limit to the number of columns tidymodels can handle?
How do I handle a situation where not all of the patterns are represented in the dataset or that the test dataset has patterns no found in the training dataset.
Is one-hot encoding the best way to go here or does tidymodels offer an alternative better suited to my use case.

Answer 1

{tidymodels} doesn't inherently have a limit on the number of columns it can handle. The bottleneck you are more likely to see would be in the parsnip engines no being able to handle many columns. {tidymodels} would also complain if you are having so many levels in the column you are applying one-hot encoding to, that the resulting data.frame doesn't fit in your RAM.
As long as the Pattern column in the data.frame you are supplying to recipe() is a factor with all the possible levels then step_dummy() will make sure all the levels will be created in both the training data set and testing data set. Since you know all the possible values Pattern can take you should manually specify the levels in the training data set you are supplying to recipe().
library(tidymodels)
patterns <- c("X", "4", "3", "2", "1", "0", "T")
patern_constructor <- function(x) {
paste(sample(patterns, size = 3, replace = FALSE, prob = seq_along(patterns)),
collapse = "")
}
all_patterns <- expand.grid(patterns, patterns, patterns) %>%
tidyr::unite(pattern, sep = "") %>%
pull(pattern)
data_patterns <- tibble(value = factor(sample(c(0,1), 10000, TRUE), c(0, 1)),
pattern = factor(map_chr(seq_len(10000), patern_constructor),
levels = all_patterns))
rec_patterns <- recipe(value ~ pattern, data = data_patterns) %>%
step_dummy(pattern, one_hot = TRUE) %>%
prep()
bake(rec_patterns, new_data = data_patterns)
#> # A tibble: 10,000 x 344
#> value pattern_XXX pattern_X4XX pattern_X3XX pattern_X2XX pattern_X1XX
#> <fct> <dbl> <dbl> <dbl> <dbl> <dbl>
#> 1 1 0 0 0 0 0
#> 2 0 0 0 0 0 0
#> 3 1 0 0 0 0 0
#> 4 1 0 0 0 0 0
#> 5 1 0 0 0 0 0
#> 6 0 0 0 0 0 0
#> 7 0 0 0 0 0 0
#> 8 1 0 0 0 0 0
#> 9 1 0 0 0 0 0
#> 10 1 0 0 0 0 0
#> # … with 9,990 more rows, and 338 more variables: pattern_X0XX <dbl>,
#> # pattern_TXX <dbl>, pattern_X4X <dbl>, pattern_X44X <dbl>,
#> # pattern_X34X <dbl>, pattern_X24X <dbl>, pattern_X14X <dbl>, …
bake(rec_patterns, new_data = tibble(value = factor(1),
pattern = factor("000")))
#> # A tibble: 1 x 344
#> value pattern_XXX pattern_X4XX pattern_X3XX pattern_X2XX pattern_X1XX
#> <fct> <dbl> <dbl> <dbl> <dbl> <dbl>
#> 1 1 0 0 0 0 0
#> # … with 338 more variables: pattern_X0XX <dbl>, pattern_TXX <dbl>,
#> # pattern_X4X <dbl>, pattern_X44X <dbl>, pattern_X34X <dbl>,
#> # pattern_X24X <dbl>, pattern_X14X <dbl>, pattern_X04X <dbl>, …
Created on 2021-03-04 by the reprex package (v0.3.0)
This will depend on the distribution of patterns you see in your data. You may have to remove patterns that didn't show up in your training data with step_zv() if the model you are using can't handle zero variance predictors. If you are having a highly skewed distribution you could use step_other() to collapse infrequent levels into one level. The embed package also includes some steps to combine multiple levels into a lower space.