13.3 Finalizing the model

Two methods:

• manually choose parameters, or

• select_best()

select_best(mlp_sfd_tune, metric = "iic")

# or, a manual selection that corresponds to the regular method

reg_param <- 
  tibble(
    num_comp = 0,
    epochs = 200,
    hidden_units = 5,
    penalty = 1   # log10
  )

# A tibble: 1 x 5
  hidden_units penalty epochs num_comp .config              
         <int>   <dbl>  <int>    <int> <chr>                
1            9 0.00780    158       14 Preprocessor06_Model1

final_sfd_wflow <- 
  mlp_wflow %>% 
  finalize_workflow(select_best(mlp_sfd_tune, metric = "rmse")) 

Now the model can fit the entire training set:

final_sfd_fit <- 
  final_sfd_wflow %>% 
  fit(Chicago_train)

This object can now be used to make future predictions on new data.

final_sfd_fit %>% 
   predict(new_data = Chicago_test) %>% 
   bind_cols(Chicago_test) %>% 
   rmse(truth = ridership, estimate = .pred)

# A tibble: 1 x 3
  .metric .estimator .estimate
  <chr>   <chr>          <dbl>
1 rmse    standard        3.01

final_sfd_fit %>% 
   predict(new_data = Chicago_test) %>% 
   bind_cols(Chicago_test) %>% 
   mutate(weekday = wday(date, 
                         label = TRUE)) %>% 
   ggplot(aes(ridership, 
              .pred, 
              color = weekday)) +
   geom_point() +
   geom_abline(color = "red") +
   scale_color_brewer(type = "qual") +
   coord_fixed() +
   expand_limits(x = 0, y = 0)

ggsave("images/13_test_performance.png")

I learned here that the recipe must exclude the dependent variable for the predict() to run on test data.