In brian-j-smith/MachineShop: Machine Learning Models and Tools

ModelSpecification Class

The r rdoc_url("ModelSpecification") class allows response and predictor variables and a model that defines the relationship between them to be packaged together into a single container for ease of model fitting and assessment. The inputs and models described previously can be included in a r rdoc_url("ModelSpecification"). Resampling control and a method for optimizing tuning parameters are also part of the specification. A r rdoc_url("ModelSpecification") object with its default or a user-specified control method has the following characteristics.

1. Tuning of input and model objects is performed simultaneously over a global grid of their parameter values.
2. The specification's control method overrides those of any included r rdoc_url("TunedInput") or r rdoc_url("TunedModel").
3. r rdoc_url("ModeledInput"), r rdoc_url("SelectedInput"), and r rdoc_url("SelectedModel") objects are not allowed in the specification.

The default tuning parameter optimization method is an exhaustive grid search. Other methods can be employed with the r rdoc_url("set_optim") functions.

| Optimization Function | Method | |:-----------------------------------|:-------------------------------------------------------------------------------| | r rdoc_url("set_optim_bayes()") | Bayesian optimization with a Gaussian process model [@snoek:2012:PBO] | | r rdoc_url("set_optim_bfgs()") | Limited-memory modification of quasi-Newton BFGS optimization [@byrd:1994:LMA] | | r rdoc_url("set_optim_grid()") | Exhaustive or random grid search | | r rdoc_url("set_optim_method()") | User-defined optimization function | | r rdoc_url("set_optim_pso()") | Particle swarm optimization [@bratton:2007:DSP] | | r rdoc_url("set_optim_sann()") | Simulated annealing [@belisle:1992:CTC] |

## Preprocessing recipe with PCA steps
pca_rec <- recipe(y ~ ., data = surv_train) %>%
  role_case(stratum = y) %>%
  step_center(all_predictors()) %>%
  step_scale(all_predictors()) %>%
  step_pca(all_predictors(), id = "PCA")

## Tuning grid of number of PCA components
pca_grid <- expand_steps(
  PCA = list(num_comp = 1:3)
)

## Model specification
(modelspec <- ModelSpecification(
  TunedInput(pca_rec, grid = pca_grid),
  TunedModel(GBMModel),
  control = surv_means_control
))

## Model fit with Bayesian optimization
bayes_fit <- modelspec %>% set_optim_bayes %>% fit
as.MLModel(bayes_fit)

Alternatively, a value of NULL may be specified as the control method in a r rdoc_url("ModelSpecifiction") so that any package input or model object is allowed, object-specific control structures and training parameters are used for selection and tuning, and objects are trained sequentially with nested resampling rather than simultaneously with a global grid. This, in essence, is how input and model objects are handled if passed separately in to the r rdoc_url("fit()") and r rdoc_url("resample()") functions apart from a r rdoc_url("ModelSpecification"). Having them together in a r rdoc_url("ModelSpecifiction") has the advantage of simplifying the function calls to be in terms of one object instead of two.

ModelSpecification(
  TunedInput(pca_rec, grid = pca_grid),
  TunedModel(GBMModel),
  control = NULL
)

brian-j-smith/MachineShop documentation built on Sept. 22, 2023, 10:01 p.m.