This package provides functions and methods to create and manipulate functions commonly used during modeling (e.g. fitting the model, making predictions, etc). It allows the user to manipulate how the same type of model can be created from different sources. It also contains a basic framework for model parameter tuning.
Modeling functions across different R packages can have very different interfaces. If you would like to try different approaches, there is a lot of syntactical minutiae to remember. The problem worsens when you move in-between platforms (e.g. doing a logistic regression in R’s
glm versus Spark’s implementation).
parsnip tries to solve this by providing similar interfaces to models. For example, if you are fitting a random forest model and would like to adjust the number of trees in the forest there are different argument names to remember:
Rather than remembering these values, a common interface to these models can be used with
The package makes the translation between
trees and the real names in each of the implementations.
ranger) but might also be methods outside of R (e.g. Stan, Spark, and others).
parsnip, similar to
recipes, separates the specification of what you want to do from the actual doing. This allows us to create broader functionality for modeling.
There are times where you would like to change a parameter from its default but you are not sure what the final value will be. This is the basis for model tuning. Since the model is not executing when created, these types of parameters can be changed using the
varying() function. This provides a simple placeholder for the value.
tune_mtry <- rand_forest(trees = 2000, mtry = varying()) tune_mtry #> Random Forest Model Specification (unknown) #> #> Main Arguments: #> mtry = varying() #> trees = 2000
This will come in handy later when we fit the model over different values of
Commonly used arguments to the modeling functions have their parameters exposed in the function. For example,
rand_forest has arguments for:
mtry: The number of predictors that will be randomly sampled at each split when creating the tree models.
trees: The number of trees contained in the ensemble.
min_n: The minimum number of data points in a node that are required for the node to be split further.
The arguments to the default function are:
args(rand_forest) #> function (mode = "unknown", mtry = NULL, trees = NULL, min_n = NULL) #> NULL
However, there might be other arguments that you would like to change or allow to vary. These are accessible using
set_engine. For example,
ranger has an option to set the internal random number seed. To set this to a specific value:
rf_with_seed <- rand_forest(trees = 2000, mtry = varying(), mode = "regression") %>% set_engine("ranger", seed = 63233) rf_with_seed #> Random Forest Model Specification (regression) #> #> Main Arguments: #> mtry = varying() #> trees = 2000 #> #> Engine-Specific Arguments: #> seed = 63233 #> #> Computational engine: ranger
To fit the model, you must:
rf_with_seed above is not ready for fitting due the
varying() parameter. We can set that parameter’s value and then create the model fit:
#> parsnip model object #> #> Ranger result #> #> Call: #> ranger::ranger(formula = formula, data = data, mtry = ~4, num.trees = ~2000, seed = sample.int(10^5, 1), num.threads = 1, verbose = FALSE) #> #> Type: Regression #> Number of trees: 2000 #> Sample size: 32 #> Number of independent variables: 10 #> Mtry: 4 #> Target node size: 5 #> Variable importance mode: none #> Splitrule: variance #> OOB prediction error (MSE): 5.57 #> R squared (OOB): 0.847
Or, using the
set.seed(56982) rf_with_seed %>% set_args(mtry = 4) %>% set_engine("randomForest") %>% fit(mpg ~ ., data = mtcars)
#> parsnip model object #> #> #> Call: #> randomForest(x = as.data.frame(x), y = y, ntree = ~2000, mtry = ~4) #> Type of random forest: regression #> Number of trees: 2000 #> No. of variables tried at each split: 4 #> #> Mean of squared residuals: 5.52 #> % Var explained: 84.3
Note that the call objects show
num.trees = ~2000. The tilde is the consequence of
parsnip using quosures to process the model specification’s arguments.
Normally, when a function is executed, the function’s arguments are immediately evaluated. In the case of
parsnip, the model specification’s arguments are not; the expression is captured along with the environment where it should be evaluated. That is what a quosure does.
parsnip uses these expressions to make a model fit call that is evaluated. The tilde in the call above reflects that the argument was captured using a quosure.