Tidymodels

Tidymodels is a collection of packages for machine learning. It is part of the tidyverse.

load packages:

library(tidymodels)

Rsample (sampling & splitting)

training- & test-split

```R data_split <- initial_split(dataset, prop=0.75, strata=target_col) training_set <- data_split %>% training() testing_set <- data_split %>% testing()

`strata` generates a stratified split with approximately the same fraction of each target class.

## Recipes (feature enigneering)

## Parsnip (model fitting)

General model formula
: ```R
outcome_variable ~ predictor_1 + predictor_2 + ... 
outcome_variable ~ . # to use all available predictors

Create model object

```R model <- linear_reg() %>% set_engine(‘lm’) %>% set_mode(‘regression’)

Fit the model
: ```R
model_fit <- model %>% 
    fit(target_col ~ ., data=train_set)

Get model summary

```R tidy(model_fit)

Returns parameter estimates, std.errors & p-values

Predict on new values
: ```R
predictions <- model_fit %>%
    predict(new_data = test_set)

Tune & Dials (hyper-)parameter optimization)

Yardstick (performance evaluation)

Requires a tibble/datset with the true and predicted outcomes.

predictions %>%
    rmse(truth = label_col, estimate = .pred) # .pred is the standard name of the prediction col

Regression quality metrics
R squared	prediction_set %>% rsq(truth = ..., estimate = ...)
Root mean squared error	... rmse()

Classification quality metrics
accuracy	prediction_set %>% accuracy(truth = ..., estimate = ...)
balanced accuracy	... bal_accuracy()
precision	... precision()
recall	... recall()
sensitivity	... sensitivity()
specificity	... specificity()
area under the curve	... roc_auc()

More metrics here.

Streamlined approach

# Fit model:
model_last_fit <- model %>% 
  last_fit(target_col ~ ., 
      split = data_split)
# Return standard quality metrics:
model_last_fit %>%
  collect_metrics() 
# Return tibble with predictions and true target values:
model_last_fit %>% 
  collect_predictions()