avoid creating regression tasks for data with categorical targets

[mlr-org]

avoid creating regression tasks for data with categorical targets

[mllg]

TaskRegr:meuse has the following unsupported feature types: factor

That means that the learner you are trying to apply does not support factor features. You must convert them to numeric first or enhance your learner with one-hot encoding using mlr3pipelines:

library(mlr3pipelines)
learner = po("encode") %>>% learner

[mohammadreza-sheykhmousa]

@giuseppec @mllg I tried to train a regression model where target is numeric but covariates are a mix of categorical and numerical var and I face the error as the following:
Error: <TaskRegr:meuse> has the following unsupported feature types: factor
Can you please help me here?

[frdanconia]

@mohammadreza-sheykhmousa @giuseppec @mllg I have the same problem, how to create regression task with numeric target (of course) and categorical covariates? I would be very grateful for your help.

[lebensterben]

1. @giuseppec There is nothing wrong to use even linear regression on classification.
2. @mohammadreza-sheykhmousa @frdanconia for a factor variable of n level, create n-1 binary variables as indicator variables

[mllg]

I have the same problem, how to create regression task with numeric target (of course) and categorical covariates? I would be very grateful for your help.

Creating such a task should be no problem, only some learners do not like the factor features. If you can give an example where the task creation fails, please reopen.

[mohammadreza-sheykhmousa]

Hi @mllg yes you're right! I solved the problem as the following:

gr = pipeline_robustify(tsk_rgr, lrn) %>>% po("learner", lrn)
      ede = resample(tsk_rgr, GraphLearner$new(gr), rsmp("holdout"))
      tsk_regr1 = ede$task$clone()

tnx to pipeline_robustify

[raff-k]

Unfortunately I have the same problem with mgcv::gam. Whereas the original model itself can process factors. The option with mlr3pipelines::po did not help me. Below is an example.

library(dplyr)
library(mlr3)
library(mlr3extralearners)
library(mlr3pipelines)
library(mgcv)

# Example from here:
# https://mlr3extralearners.mlr-org.com/reference/mlr_learners_classif.gam.html

# ... get data
t <- mlr3::tsk("spam")
t_data <- t$data()

# ... (re-)create the task
t_re <- as_task_classif( 
  id = "spam", 
  target = "type", 
  positive = "spam",
  x = t_data
)

# ... init mgcv gam learner
l <- mlr3::lrn("classif.gam",
              formula =  type ~ s(george) + s(charDollar) + s(edu) + ti(george, edu))

# ... train and get gam
l$train(t_re)
l$model # ... creates some output -> example of mlr3-page successfully reproduced


# NOW, the error due to the created factor
set.seed(123)
t_data_fac <- t_data %>%
        dplyr::mutate(fac = sample(x = c(1:3), size = nrow(t_data), replace = T) %>% as.factor())

# create task with additional factor variable
t_fac = as_task_classif( 
  id = "spam_fac", 
  target = "type", 
  positive = "spam",
  x = t_data_fac
)

# ... init mgcv gam learner with factor
l_fac <- mlr3::lrn("classif.gam",
              formula =  type ~ s(george) + s(charDollar) + s(edu) + ti(george, edu) + fac)


# ... train and get gam
l_fac$train(t_fac)

# Error: <TaskClassif:spam_fac> has the following unsupported feature types: factor

# Here is the page of mlr3-implementation of mgcv::gam
# https://mlr3extralearners.mlr-org.com/reference/mlr_learners_classif.gam.html
# Perhaps in Feature Types: “logical”, “integer”, “numeric” --> "factor" is missing?

# ... but normally, factors are no issues for mgcv::gam
l_gam <- mgcv::gam(formula = type ~ s(george) + s(charDollar) + s(edu) + ti(george, edu) + fac, 
                 data = t_data_fac, family = "binomial")

l_gam %>% summary()

# Parametric coefficients:
# Estimate Std. Error z value Pr(>|z|)
# (Intercept) 2.441e+03  2.479e+05   0.010    0.992
# fac2        4.675e-02  9.814e-02   0.476    0.634
# fac3        5.879e-02  9.689e-02   0.607    0.544



# Now, trying mlr3pipelines solution, but it does not work
l_fac_enc <- mlr3pipelines::po("encode")  %>>% 
  mlr3::lrn("classif.gam", formula =  type ~ s(george) + s(charDollar) + s(edu) + ti(george, edu) + fac)

l_fac_enc$train(input = t_fac)
# Error in eval(predvars, data, env) : object 'fac' not found
# This happened PipeOp classif.gam's $train()