Support multiple interfaces of (not to) an engine

[hfrick]

We currently only allow one interface of an engine, set by set_fit(). Some engines have multiple interfaces themselves but we don't leverage that. This SO post runs into troubles with the formula interface of kernlab::ksvm() which could be resolved by using the matrix interface of the kernlab function. The workflow does use the tidymodels matrix interface but eventually translates it to the formula interface of kernlab because that's how it's registered in parnsip.

This single translation point from parsnip to engine is also a challenge for https://github.com/tidymodels/censored/issues/311

library(tidymodels)
library(kernlab)
# [...]

x <- matrix(rnorm(2000000), nrow = 100, ncol = 20000)
colnames(x) <- paste0("x", 1:20000)
y <- rnorm(n = 100)
data <- cbind(y, x) %>% as.data.frame()

# formula interface struggles
svm.train <- ksvm(y ~ ., type="eps-svr", data = data, kernel ="rbfdot")
#> Error: protect(): protection stack overflow

# matrix interface works
svm.train <- ksvm(x = x, y = y, type = "eps-svr", kernel ="rbfdot")

# tidymodels always uses the formula interface of kernlab itself, 
# regardless of the tidymodels interface

svm_spec <- svm_rbf(engine = "kernlab", mode = "regression")

fit_f <- fit(svm_spec, y ~ ., data = data)
#> Error: protect(): protection stack overflow
fit_xy <- fit_xy(svm_spec, x = x, y = y)
#> Error: protect(): protection stack overflow

^{Created on 2024-04-23 with reprex v2.1.0}

[simonpcouch]

Do you anticipate any downsides to, instead, just fully switching to registering kernlab::ksvm() via its XY interface? Same question would go for coxnet_train() as well, I guess.

[hfrick]

I'm not familiar with kernlab so can't give a qualified answer on that right now :)

For glmnet/coxnet: 😬

There is a fundamental design clash wrt to stratification. glmnet expects the response to be stratified which would mean that we would not have stratification information available at prediction time with tidymodels. To get out of that, coxnet_train() handles the translation of stratification.

[EmilHvitfeldt]

having multiple interfaces might be nice once we have sparse tibble support. all sparsity should be done using _xy, but a given model might perform better non-sparse using a formula interface

[hfrick]

+1 on the sparsity comment - one example (of probably a few more?) is https://github.com/tidymodels/censored/issues/276