step_o2pls() creates a specification of a recipe step that removes
num_comp outcome-orthogonal components from the selected predictors
with o2pls(). Unlike the other orthogonalization steps, the filter can
be estimated against several outcomes at once.
Arguments
- recipe
A recipe object. The step will be added to the sequence of operations for this recipe.
- ...
One or more selector functions to choose the predictors (for example, the spectral channels). See
recipes::selections().- role
Not used by this step, since no new variables are created.
- trained
A logical indicating whether the step has been trained.
- outcome
The outcome variables, as bare names or selectors. If
NULL(default), all the outcomes of the recipe are used.- num_comp
The number of outcome-orthogonal components to remove (
nxino2pls()).- joint_comp
The number of joint (predictive) components (
ncompino2pls()), at most the number of outcomes. Default is 1.- options
A list of further arguments passed to
o2pls():centerandscale.- res
The fitted filter, stored once the step has been trained.
- columns
The names of the selected predictors, stored once the step has been trained.
- skip
A logical. Should the step be skipped when the recipe is baked by
recipes::bake()? Keep the defaultFALSE, so that new data are corrected with the same filter as the training data.- id
A character string that is unique to this step.
Details
The joint (predictive) directions are the joint_comp dominant singular
vectors of \(\textbf{Y}^T\textbf{X}\), and the removed components are the
systematic variation of the predictors orthogonal to them (Trygg and Wold,
2003). Only the predictors are filtered: the outcomes are left unchanged,
so that models are fitted to, and assessed on, the measured values. (The
outcome-side filtering of o2pls(), its ny argument, does not change the
filtered predictors.)
With a single outcome and joint_comp = 1, the filtered data are those of
step_opls() and step_projected_osc() with the same num_comp.
The selected columns are replaced by the filtered values, which are
centered (and scaled, if options = list(scale = TRUE)). The filter uses
the outcomes, so it is estimated on the training data only (on the
analysis set of each resample within tune::tune_grid()); the outcomes are
not needed when new data are baked.
Several outcomes
Declare the outcomes in the recipe formula; a model that handles several
outcomes, such as parsnip::pls() with the mixOmics engine, then predicts
them all, with one column per outcome (.pred_K, .pred_Ca, ...):
rec <- recipe(K + Ca ~ ., data = spectra) |>
step_o2pls(all_predictors(), num_comp = 2, joint_comp = 2)
model <- parsnip::pls(num_comp = 2) |>
set_mode("regression") |>
set_engine("mixOmics", scale = FALSE)
fitted <- fit(workflow(rec, model), data = spectra)
predict(fitted, new_data = new_spectra)Use outcome to estimate the filter against some of the outcomes only.
Tuning
num_comp and joint_comp can be tuned with tune::tune() when the
recipe has a single outcome. Their default ranges are dials::num_comp()
with values 1 to 4 and 1 to 3; joint_comp cannot exceed the number of
outcomes.
tune::tune_grid() does not support several outcomes. Instead, create
one workflow per outcome with the workflowsets package, and tune them all
with workflowsets::workflow_map("tune_grid", ...). Each recipe has a
single outcome, for the model, but the filter can still be estimated
against all the outcomes: give the others a role of their own
("reference" here), not needed to bake new data, and select them with
outcome. Each outcome gets its own number of filter and model components:
outcomes <- c("K", "Ca", "Mg")
recipes <- purrr::map(outcomes, \(y) {
recipe(reformulate(".", response = y), data = spectra) |>
update_role(all_of(setdiff(outcomes, y)), new_role = "reference") |>
update_role_requirements("reference", bake = FALSE) |>
step_o2pls(all_predictors(), outcome = c(all_outcomes(), has_role("reference")),
num_comp = tune("filter"), joint_comp = length(outcomes))
}) |>
purrr::set_names(outcomes)
model <- parsnip::pls(num_comp = tune()) |>
set_mode("regression") |>
set_engine("mixOmics", scale = FALSE)
wf_set <- workflow_set(preproc = recipes, models = list(pls = model))
res <- workflow_map(wf_set, "tune_grid", resamples = vfold_cv(spectra, v = 5),
grid = 10, metrics = metric_set(rmse), seed = 1)
# best settings and final model of each outcome
fits <- purrr::map(purrr::set_names(res$wflow_id), \(id) {
best <- select_best(extract_workflow_set_result(res, id), metric = "rmse")
extract_workflow(res, id) |>
finalize_workflow(best) |>
fit(data = spectra)
})
purrr::map(fits, \(f) predict(f, new_data = new_spectra))Select the outcomes by role (all_outcomes(), has_role()) rather than
with all_of(outcomes), which refers to a variable outside the recipe.
To filter each outcome against itself only, drop the other outcomes from
its recipe instead; the step is then equivalent to step_opls(). Read
the results outcome by outcome: workflowsets::rank_results() would rank
workflows of different outcomes, whose RMSE are in different units.
Tidying
tidy() returns a tibble with columns terms (the
selected predictors), num_comp, joint_comp and id.
References
Trygg, J., Wold, S., (2003). O2-PLS, a two-block (X–Y) latent variable regression (LVR) method with an integral OSC filter. J. Chemom. 17(1):53–64.
Examples
data(forageLIBS)
# potassium, calcium and the spectral channels
dat <- forageLIBS[-c(1:2, 4:10, 12:14)]
rec <- recipes::recipe(K + Ca ~ ., data = dat[1:300, ]) |>
step_o2pls(recipes::all_predictors(), num_comp = 2, joint_comp = 2)
prepped <- recipes::prep(rec)
recipes::tidy(prepped, number = 1)
#> # A tibble: 7,152 × 4
#> terms num_comp joint_comp id
#> <chr> <dbl> <dbl> <chr>
#> 1 199.3771616 2 2 o2pls_6YGuu
#> 2 199.4644141 2 2 o2pls_6YGuu
#> 3 199.5516666 2 2 o2pls_6YGuu
#> 4 199.6389192 2 2 o2pls_6YGuu
#> 5 199.7261717 2 2 o2pls_6YGuu
#> 6 199.8134242 2 2 o2pls_6YGuu
#> 7 199.9006767 2 2 o2pls_6YGuu
#> 8 199.9879292 2 2 o2pls_6YGuu
#> 9 200.0751817 2 2 o2pls_6YGuu
#> 10 200.1624343 2 2 o2pls_6YGuu
#> # ℹ 7,142 more rows
dim(recipes::bake(prepped, new_data = dat[301:368, ]))
#> [1] 68 7154