6.1 Create a Model

6.1.1 Different Model Interfaces

Different-interfaces

• Model Interfaces
  • Different Implementations = Different Interfaces
  • Linear Regression can be implemented in many ways
    • Ordinary Least Squares
    • Regularized Linear Regression
    • …

• {stats}
  • takes formula
  • uses data.frame

lm(formula, data, ...)

• {glmnet}
  • Has x/y interface
  • Uses a matrix

glmnet(x = matrix, y = vector, family = "gaussian", ...)

6.1.2 Model Specification

model specification

• {tidymodels}/{parsnip} - Philosophy is to unify & make interfaces more predictable.
  • Specify model type (e.g. linear regression, random forest …)
    • linear_reg()
    • rand_forest()
  • Specify engine (i.e. package implementation of algorithm)
    • set_engine("some package's implementation")
  • declare mode (e.g. classification vs linear regression)
    • use this when model can do both classification & regression
    • set_mode("regression")
    • set_mode("classification")

• Bringing it all together

lm_model_spec <- 
  linear_reg() %>% # specify model
  set_engine("lm") # set engine


lm_model_spec

## Linear Regression Model Specification (regression)
## 
## Computational engine: lm

6.1.3 Model Fitting

From above we will use our existing model specification

• fit()
  • any nominal or categorical variables will be split out into dummy columns
  • most formula methods also turn do the same thing
• fit_xy
  • delays creating dummy variable and has underlying model function

# create model fit using formula
lm_form_fit <- 
  lm_model_spec %>% 
  fit(Sale_Price ~ Longitude + Latitude, data = ames_train)


# create model fit using x/y
lm_xy_fit <- 
  lm_model_spec %>% 
  fit_xy(
    x = ames_train %>% select(Longitude, Latitude),
    y = ames_train %>% pull(Sale_Price)
    )

6.1.4 Generalized Model Arguments

• Like the varying interfaces, model parameters differ from implementation to implementation
• two level of model arguments
  • main arguments - Parameters aligned with the mathematical vehicle
  • engine arguments - Parameters aligned with the package implementation of the mathematical algorithm

argument	ranger	randomForest	sparklyr
sampled predictors	mtry	mtry	feature_subset_strategy
trees	num.tress	ntree	num_trees
data points to split	min.node.size	nodesize	min_instances_per_node

argument	parsnip
sampled predictors	mtry
trees	trees
data points to split	min_n

Parsnip in Action

• The translate() provides the mapping from the parsnips interface to the each individual package’s implementation of the algorithm.

# stats implementation
linear_reg() %>% 
  set_engine("lm") %>% 
  translate()

## Linear Regression Model Specification (regression)
## 
## Computational engine: lm 
## 
## Model fit template:
## stats::lm(formula = missing_arg(), data = missing_arg(), weights = missing_arg())

# glmnet implementation
linear_reg(penalty = 1) %>% 
  set_engine("glmnet") %>% 
  translate()

## Linear Regression Model Specification (regression)
## 
## Main Arguments:
##   penalty = 1
## 
## Computational engine: glmnet 
## 
## Model fit template:
## glmnet::glmnet(x = missing_arg(), y = missing_arg(), weights = missing_arg(), 
##     family = "gaussian")