Go through all registered models and identify if they accepts sparse data

[EmilHvitfeldt]

Basically, for all models that has allow_sparse_x = FALSE, douple-check if that is still the case. If any are found, then they should be tested as part of #1220

should be done for all extension models too

[EmilHvitfeldt]

parsnip_pkgs <- revdepcheck::cran_revdeps(
  "parsnip",
  dependencies = c("Depends", "Imports")
)
parsnip_pkgs <- c(parsnip_pkgs, "parsnip")
# These ignore the tidymodels design principles and/or don't work with the
# broader ecosystem or we don't don't have any models in them
excl <- c(
  "additive",
  "bayesian",
  "cuda.ml",
  "SSLR",
  "workflowsets",
  "workflows",
  "tune",
  "tidymodels",
  "shinymodels",
  "stacks"
)
parsnip_pkgs <- parsnip_pkgs[!(parsnip_pkgs %in% excl)]

pak::pak(parsnip_pkgs)

# Load them then get the model data base
loaded <- purrr::map_lgl(
  parsnip_pkgs,
  ~ suppressPackageStartupMessages(require(.x, character.only = TRUE))
)
table(loaded)

# h2o overwrites soooo many functions; this may take a few minutes
conflicted::conflict_prefer_all("base", loser = "h2o", quiet = TRUE)

origin_fit <- rlang::env_get_list(
  env = parsnip::get_model_env(),
  nms = ls(parsnip::get_model_env(), pattern = "_fit$")
) |>
  purrr::list_rbind(names_to = "thing")

origin_pkg <- rlang::env_get_list(
  env = parsnip::get_model_env(),
  nms = ls(parsnip::get_model_env(), pattern = "_pkgs$")
) |>
  purrr::list_rbind(names_to = "thing")

library(tidyverse)

all_fit_calls <- origin_fit |>
  slice(-c(121, 122)) |>
  mutate(
    func = map(value, "func"),
    func_fun = map_chr(func, "fun"),
    func_pkg = map_chr(func, \(x) if_else(is.null(x["pkg"]), NA, x["pkg"]))
  )

all_fit_calls |>
  dplyr::filter(!is.na(func_pkg)) |> # All timeseries stuff
  mutate(exp = glue::glue("{func_pkg}::{func_fun}")) |>
  distinct(exp) |>
  pull(exp) |> 
  sort()

Result:

aorsf::orsf
baguette::bagger
bonsai::train_lightgbm
brulee::brulee_linear_reg
brulee::brulee_logistic_reg
brulee::brulee_mlp
brulee::brulee_mlp_two_layer
brulee::brulee_multinomial_reg
card::cosinor
censored::blackboost_train
censored::coxnet_train
dbarts::bart
discrim::fit_regularized_linear
discrim::fit_regularized_quad
discrim::klar_bayes_wrapper
earth::earth
flexsurv::flexsurvreg
flexsurv::flexsurvspline
flexsurvcure::flexsurvcure
glmnet::glmnet
ipred::bagging
kernlab::ksvm
kknn::train.kknn
klaR::rda
LiblineaR::LiblineaR
liquidSVM::svm
MASS::lda
MASS::qda
mda::fda
mgcv::gam
naivebayes::naive_bayes
nnet::multinom
nnet::nnet
offsetreg::glm_offset
offsetreg::glmnet_offset
offsetreg::rpart_exposure
offsetreg::xgb_train_offset
parsnip::C5.0_train
parsnip::cforest_train
parsnip::ctree_train
parsnip::keras_mlp
parsnip::nullmodel
parsnip::xgb_train
pec::pecRpart
plsmod::pls_fit
pscl::hurdle
pscl::zeroinfl
quantreg::rq
randomForest::randomForest
ranger::ranger
rpart::rpart
rstanarm::stan_glm
rules::c5_fit
rules::cubist_fit
rules::xrf_fit
sda::sda
sparklyr::ml_decision_tree_classifier
sparklyr::ml_decision_tree_regressor
sparklyr::ml_gradient_boosted_trees
sparklyr::ml_linear_regression
sparklyr::ml_logistic_regression
sparklyr::ml_random_forest
stats::glm
stats::lm
survival::coxph
survival::survreg
tidysdm::maxnet_fit

[EmilHvitfeldt]

Going through, these are all the models that can benefit

# A tibble: 16 × 3
   thing                    engine    mode               
   <chr>                    <chr>     <chr>              
 1 boost_tree           xgboost   regression         
 2 boost_tree           xgboost   classification     
 3 boost_tree           lightgbm  regression         
 4 boost_tree           lightgbm  classification     
 5 linear_reg           glmnet    regression         
 6 logistic_reg         glmnet    classification     
 7 logistic_reg         LiblineaR classification     
 8 multinom_reg         glmnet    classification     
 9 null_model           parsnip   regression         
10 null_model           parsnip   classification     
11 poisson_reg          glmnet    regression         
12 proportional_hazards glmnet    censored regression
13 rand_forest          ranger    classification     
14 rand_forest          ranger    regression         
15 svm_linear           LiblineaR regression         
16 svm_linear           LiblineaR classification     

I need to look into

• boost_tree for lightgbm
• proportional_hazards for glmnet
• null_model for parsnip

[EmilHvitfeldt]

proportional_hazards() isn't gonna be possible at the moment, because coxnet_train() uses formula interface, which doesn't work with sparsity

[EmilHvitfeldt]

Listed things in

• make sure lightgbm works with sparse data bonsai#100
• make null_model() work with sparse data #1265

[github-actions]

This issue has been automatically locked. If you believe you have found a related problem, please file a new issue (with a reprex: https://reprex.tidyverse.org) and link to this issue.