Add argument for one hot encoding to parsnip

[juliasilge]

This PR connects tidymodels/hardhat#140 to parsnip, with one_hot as an encoding option in registering a model.

library(parsnip)

c("boost_tree",
  "decision_tree",
  "linear_reg",
  "logistic_reg",
  "mars",
  "mlp",
  "multinom_reg",
  "nearest_neighbor",
  "null_model",
  "rand_forest",
  "surv_reg",
  "svm_poly",
  "svm_rbf") %>%
  purrr::map_dfr(get_encoding) %>%
  knitr::kable()

model	engine	mode	predictor_indicators	one_hot
boost_tree	xgboost	regression	TRUE	FALSE
boost_tree	xgboost	classification	TRUE	FALSE
boost_tree	C5.0	classification	FALSE	FALSE
boost_tree	spark	regression	TRUE	FALSE
boost_tree	spark	classification	TRUE	FALSE
decision_tree	rpart	regression	FALSE	FALSE
decision_tree	rpart	classification	FALSE	FALSE
decision_tree	C5.0	classification	FALSE	FALSE
decision_tree	spark	regression	TRUE	FALSE
decision_tree	spark	classification	TRUE	FALSE
linear_reg	lm	regression	TRUE	FALSE
linear_reg	glmnet	regression	TRUE	TRUE
linear_reg	stan	regression	TRUE	FALSE
linear_reg	spark	regression	TRUE	FALSE
linear_reg	keras	regression	TRUE	FALSE
logistic_reg	glm	classification	TRUE	FALSE
logistic_reg	glmnet	classification	TRUE	TRUE
logistic_reg	spark	classification	TRUE	FALSE
logistic_reg	keras	classification	TRUE	FALSE
logistic_reg	stan	classification	TRUE	FALSE
mars	earth	regression	FALSE	FALSE
mars	earth	classification	FALSE	FALSE
mlp	keras	regression	TRUE	FALSE
mlp	keras	classification	TRUE	FALSE
mlp	nnet	regression	TRUE	FALSE
mlp	nnet	classification	TRUE	FALSE
multinom_reg	glmnet	classification	TRUE	TRUE
multinom_reg	spark	classification	TRUE	FALSE
multinom_reg	keras	classification	TRUE	FALSE
multinom_reg	nnet	classification	TRUE	FALSE
nearest_neighbor	kknn	regression	TRUE	FALSE
nearest_neighbor	kknn	classification	TRUE	FALSE
null_model	parsnip	regression	FALSE	FALSE
null_model	parsnip	classification	FALSE	FALSE
rand_forest	ranger	classification	FALSE	FALSE
rand_forest	ranger	regression	FALSE	FALSE
rand_forest	randomForest	classification	FALSE	FALSE
rand_forest	randomForest	regression	FALSE	FALSE
rand_forest	spark	classification	TRUE	FALSE
rand_forest	spark	regression	TRUE	FALSE
surv_reg	flexsurv	regression	TRUE	FALSE
surv_reg	survival	regression	TRUE	FALSE
svm_poly	kernlab	regression	FALSE	FALSE
svm_poly	kernlab	classification	FALSE	FALSE
svm_rbf	kernlab	regression	FALSE	FALSE
svm_rbf	kernlab	classification	FALSE	FALSE
svm_rbf	liquidSVM	regression	FALSE	FALSE
svm_rbf	liquidSVM	classification	FALSE	FALSE

^{Created on 2020-06-18 by the reprex package (v0.3.0.9001)}

I have not yet worked on any changes to form_xy() or convert_form_to_xy_fit() to use the one_hot argument. I believe convert_form_to_xy_fit() will need to use some of the new hardhat work, like contrasts, etc.

[juliasilge]

Ah actually, I just realized I may have our plan not laid out in my head clearly. parsnip does not depend on hardhat at all right now. Let's chat more about the plan for, for example, contrasts.

[juliasilge]

Closes #326

[topepo]

After doing some tests, I think that the glmnet models should be changed to the traditional indicator scheme.

My recollection about this was half right - while an intercept is not created with the other coefficients (or regularized), one is calculated after parameter estimation.

I did some sanity checking with a simple model containing a single factor predictor. Doing a one-hot encoding would results in incorrect and inaccurate parameters (unless parsnip were to choose the glmnet option to not estimate the intercept and that's too much of a deviation).

I"ll update the PR to change those back.

Also, I thought that we were going to do one-hot for xgboost. Am I mis-remembering that?

[juliasilge]

The latest version here only handles the "traditional" indicators when going to convert_form_to_xy_fit(), via:

indicators <- indicators == "traditional"

I'm leaving this as draft because we need to still handle the one hot case, with different contrasts and all that.

The current results of get_encoding() are:

library(parsnip)

c("boost_tree",
  "decision_tree",
  "linear_reg",
  "logistic_reg",
  "mars",
  "mlp",
  "multinom_reg",
  "nearest_neighbor",
  "null_model",
  "rand_forest",
  "surv_reg",
  "svm_poly",
  "svm_rbf") %>%
  purrr::map_dfr(get_encoding) %>%
  knitr::kable()

model	engine	mode	predictor_indicators
boost_tree	xgboost	regression	one_hot
boost_tree	xgboost	classification	one_hot
boost_tree	C5.0	classification	none
boost_tree	spark	regression	traditional
boost_tree	spark	classification	traditional
decision_tree	rpart	regression	none
decision_tree	rpart	classification	none
decision_tree	C5.0	classification	none
decision_tree	spark	regression	traditional
decision_tree	spark	classification	traditional
linear_reg	lm	regression	traditional
linear_reg	glmnet	regression	traditional
linear_reg	stan	regression	traditional
linear_reg	spark	regression	traditional
linear_reg	keras	regression	traditional
logistic_reg	glm	classification	traditional
logistic_reg	glmnet	classification	traditional
logistic_reg	spark	classification	traditional
logistic_reg	keras	classification	traditional
logistic_reg	stan	classification	traditional
mars	earth	regression	none
mars	earth	classification	none
mlp	keras	regression	traditional
mlp	keras	classification	traditional
mlp	nnet	regression	traditional
mlp	nnet	classification	traditional
multinom_reg	glmnet	classification	traditional
multinom_reg	spark	classification	traditional
multinom_reg	keras	classification	traditional
multinom_reg	nnet	classification	traditional
nearest_neighbor	kknn	regression	traditional
nearest_neighbor	kknn	classification	traditional
null_model	parsnip	regression	none
null_model	parsnip	classification	none
rand_forest	ranger	classification	none
rand_forest	ranger	regression	none
rand_forest	randomForest	classification	none
rand_forest	randomForest	regression	none
rand_forest	spark	classification	traditional
rand_forest	spark	regression	traditional
surv_reg	flexsurv	regression	traditional
surv_reg	survival	regression	traditional
svm_poly	kernlab	regression	none
svm_poly	kernlab	classification	none
svm_rbf	kernlab	regression	none
svm_rbf	kernlab	classification	none
svm_rbf	liquidSVM	regression	none
svm_rbf	liquidSVM	classification	none

^{Created on 2020-06-26 by the reprex package (v0.3.0.9001)}

[github-actions]

This pull request 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.