Optimize the boosted ensemble

Machine Learning with Tree-Based Models in R

Sandro Raabe

Data Scientist

Starting point: untuned performance

collect_metrics(cv_results)

# A tibble: 1 x 3
  .metric   .mean       n   
  <chr>     <dbl>   <int>             
1 roc_auc   0.951       5

95% - not bad for untuned model!

Tuning workflow

tune function

Tuning workflow

grid generation

Tuning workflow

cross-validation

Tuning workflow

tune_grid function

Tuning workflow

select_best function

Tuning workflow

tuning workflow

Tuning workflow

final fit function

Step 1: Create the tuning spec

# Create the specification with placeholders
boost_spec <- boost_tree(

    trees = 500,
    learn_rate = tune(),
    tree_depth = tune(),
    sample_size = tune()) %>%

  set_mode("classification") %>%
  set_engine("xgboost")

Boosted Tree Model Specification (classification)

Main Arguments:
  trees = 500
  tree_depth = tune()
  learn_rate = tune()
  sample_size = tune()

Computational engine: xgboost

Usual specification
Major difference: use tune() to create placeholders for values to be tuned

Step 2: Create the tuning grid

# Create a regular grid
tunegrid_boost <- grid_regular(parameters(boost_spec), 

                               levels = 2)

# A tibble: 8 x 3
  tree_depth     learn_rate sample_size
       <int>          <dbl>       <dbl>
1          1   0.0000000001         0.1
2         15   0.0000000001         0.1
3          1   0.1                  0.1
4         15   0.1                  0.1
5          1   0.0000000001         1  
6         15   0.0000000001         1  
7          1   0.1                  1  
8         15   0.1                  1

# Create a random grid
grid_random(parameters(boost_spec),

            size = 8)

# A tibble: 8 x 3
  tree_depth    learn_rate   sample_size
       <int>         <dbl>         <dbl>
1         11   0.0000000249        0.858
2         12   0.00000000392       0.856
3         15   0.0000000131        0.220
4         15   0.0000216           0.125
5         10   0.00000000537       0.759
6         14   0.0395              0.270
7          2   0.000000828         0.904
8          9   0.0000254           0.473

Step 3: The tuning

Arguments for tune_grid():

Dummy specification
Model formula
Resamples/folds
Parameter grid
Metric list in metric_set()

Function call:

# Tune along the grid
tune_results <- tune_grid(

    boost_spec,

    still_customer ~ .,

    resamples = vfold_cv(customers_train, v = 6),

    grid = tunegrid_boost,

    metrics = metric_set(roc_auc))

Visualize the result

# Plot the results
autoplot(tune_results)

tuning results

Step 4: Finalize the model

# Select the final hyperparameters
best_params <- select_best(tune_results)

best_params

# A tibble: 1 x 4
  tree_depth learn_rate sample_size .config
       <int>      <dbl>       <dbl> <chr> 
1          8        0.1        0.55 Model17

# Finalize the specification
final_spec <- finalize_model(boost_spec, 
                             best_params)
final_spec

Boosted Tree Model Specification

Main Arguments:
  trees = 500
  tree_depth = 8
  learn_rate = 0.1
  sample_size = 0.55

Computational engine: xgboost

Last step: Train the final model

final_model <- final_spec %>% fit(formula = still_customer ~ .,
                                  data = customers_train)

final_model

Fit time:  2.3s 
##### xgb.Booster
raw: 343.8 Kb 
nfeatures : 37 
evaluation_log:
    iter training_error
       1       0.046403                  
     100       0.002592

Your turn!

Machine Learning with Tree-Based Models in R