Measuring cross-validation performance

Machine Learning in the Tidyverse

Dmitriy (Dima) Gorenshteyn

Lead Data Scientist, Memorial Sloan Kettering Cancer Center

Measuring Performance

Measuring Performance - Truth

Measuring Performance - Prediction

Measuring Performance

Mean Absolute Error

Ingredients for Performance Measurement

1) Actual life_expectancy values
2) Predicted life_expectancy values
3) A metric to compare 1) & 2)

1) Extract the actual values

cv_prep_lm <- cv_models_lm %>% 
  mutate(validate_actual = map(validate, ~.x$life_expectancy))

The predict() & map2() functions

predict(model, data)

map2(.x = model, .y = data, .f = ~predict(.x, .y))

2) Prepare the predicted values

cv_prep_lm <- cv_eval_lm %>% 
  mutate(validate_actual = map(validate, ~.x$life_expectancy),
         validate_predicted = map2(model, validate, ~predict(.x, .y)))

3) Calculate MAE

library(Metrics)
cv_eval_lm <- cv_prep_lm %>% 
  mutate(validate_mae = map2_dbl(validate_actual, validate_predicted, 
                                ~mae(actual = .x, predicted = .y)))

cv_eval_lm

#  5-fold cross-validation 
# A tibble: 5 x 8
splits       id    train validate model validate_a. validate_p validate_mae
<S3: rsplit> Fold1 <tib. <tib.   <S3.   <dbl.       <dbl.       1.47
<S3: rsplit> Fold2 <tib. <tib.   <S3.   <dbl.       <dbl.       1.51
<S3: rsplit> Fold3 <tib. <tib.   <S3.   <dbl.       <dbl.       1.44
<S3: rsplit> Fold4 <tib. <tib.   <S3.   <dbl.       <dbl.       1.48
<S3: rsplit> Fold5 <tib. <tib.   <S3.   <dbl.       <dbl.       1.68

Let's practice!

Machine Learning in the Tidyverse