MultiLR: Predictive models (cont.)

STA 210 - Spring 2022

Author

Dr. Mine Çetinkaya-Rundel

Welcome

Topics

  • Unbalanced data
  • Choosing the “final” model

Computational setup

# load packages
library(tidyverse)
library(tidymodels)
library(knitr)
library(colorblindr)
library(themis)

# set default theme and larger font size for ggplot2
ggplot2::theme_set(ggplot2::theme_minimal(base_size = 16))

From last time…

Volcanoes

The data come from The Smithsonian Institution, via TidyTuesday.

volcano <- read_csv(here::here("slides", "data/volcano.csv"))
names(volcano)
 [1] "volcano_number"           "volcano_name"            
 [3] "primary_volcano_type"     "last_eruption_year"      
 [5] "country"                  "region"                  
 [7] "subregion"                "latitude"                
 [9] "longitude"                "elevation"               
[11] "tectonic_settings"        "evidence_category"       
[13] "major_rock_1"             "major_rock_2"            
[15] "major_rock_3"             "major_rock_4"            
[17] "major_rock_5"             "minor_rock_1"            
[19] "minor_rock_2"             "minor_rock_3"            
[21] "minor_rock_4"             "minor_rock_5"            
[23] "population_within_5_km"   "population_within_10_km" 
[25] "population_within_30_km"  "population_within_100_km"

Data prep

volcano <- volcano %>%
  mutate(
    volcano_type = case_when(
      str_detect(primary_volcano_type, "Stratovolcano") ~ "Stratovolcano",
      str_detect(primary_volcano_type, "Shield") ~ "Shield",
      TRUE ~ "Other"
    ),
    volcano_type = fct_relevel(volcano_type, "Stratovolcano", "Shield", "Other")
  ) %>%
  select(
    volcano_type, latitude, longitude, 
    elevation, tectonic_settings, major_rock_1
    ) %>%
  mutate(across(where(is.character), as_factor))

Split into testing/training

set.seed(1234)

volcano_split <- initial_split(volcano)
volcano_train <- training(volcano_split)
volcano_test  <- testing(volcano_split)

Specify a model

volcano_spec <- multinom_reg() %>%
  set_engine("nnet")

volcano_spec
Multinomial Regression Model Specification (classification)

Computational engine: nnet 

Create cross validation folds

set.seed(9876)

volcano_folds <- vfold_cv(volcano_train, v = 5)
volcano_folds
#  5-fold cross-validation 
# A tibble: 5 × 2
  splits            id   
  <list>            <chr>
1 <split [574/144]> Fold1
2 <split [574/144]> Fold2
3 <split [574/144]> Fold3
4 <split [575/143]> Fold4
5 <split [575/143]> Fold5

Unbalanced data

Unbalanced data

Remember that the observed volcano types are unbalanced:

volcano %>% 
  count(volcano_type)
# A tibble: 3 × 2
  volcano_type      n
  <fct>         <int>
1 Stratovolcano   461
2 Shield          118
3 Other           379

Addressing unbalance

To address class unbalance, we generally use

  • oversampling data from levels that are less prevalent in the data
  • downsampling data from levels that are more prevalent in the data
    • e.g., step_downsample(): Removes rows of a data set to make the occurrence of levels in a specific factor level equal.

New recipe - oversample

volcano_rec3 <- recipe(volcano_type ~ ., data = volcano_train) %>%
  step_other(tectonic_settings) %>%
  step_other(major_rock_1) %>%
  step_dummy(all_nominal_predictors()) %>%
  step_zv(all_predictors()) %>%
  step_center(all_predictors()) %>%
  step_smote(volcano_type)

New recipe - downsample

volcano_rec4 <- recipe(volcano_type ~ ., data = volcano_train) %>%
  step_other(tectonic_settings) %>%
  step_other(major_rock_1) %>%
  step_dummy(all_nominal_predictors()) %>%
  step_zv(all_predictors()) %>%
  step_center(all_predictors()) %>%
  step_downsample(volcano_type)

New workflows

volcano_wflow3 <- workflow() %>%
  add_recipe(volcano_rec3) %>%
  add_model(volcano_spec)

volcano_wflow4 <- workflow() %>%
  add_recipe(volcano_rec4) %>%
  add_model(volcano_spec)

Fit resamples

volcano_fit_rs3 <- volcano_wflow3 %>%
  fit_resamples(
    volcano_folds, 
    control = control_resamples(save_pred = TRUE)
    )

volcano_fit_rs4 <- volcano_wflow4 %>%
  fit_resamples(
    volcano_folds, 
    control = control_resamples(save_pred = TRUE)
    )

Collect metrics

collect_metrics(volcano_fit_rs3)
# A tibble: 2 × 6
  .metric  .estimator  mean     n std_err .config             
  <chr>    <chr>      <dbl> <int>   <dbl> <chr>               
1 accuracy multiclass 0.517     5  0.0154 Preprocessor1_Model1
2 roc_auc  hand_till  0.693     5  0.0270 Preprocessor1_Model1
collect_metrics(volcano_fit_rs4)
# A tibble: 2 × 6
  .metric  .estimator  mean     n std_err .config             
  <chr>    <chr>      <dbl> <int>   <dbl> <chr>               
1 accuracy multiclass 0.485     5  0.0123 Preprocessor1_Model1
2 roc_auc  hand_till  0.675     5  0.0219 Preprocessor1_Model1

ROC curves - oversampling

volcano_fit_rs3 %>%
  collect_predictions() %>%
  group_by(id) %>%
  roc_curve(
    truth = volcano_type,
    .pred_Stratovolcano:.pred_Other
  ) %>%
  autoplot()

ROC curves - downsampling

volcano_fit_rs4 %>%
  collect_predictions() %>%
  group_by(id) %>%
  roc_curve(
    truth = volcano_type,
    .pred_Stratovolcano:.pred_Other
  ) %>%
  autoplot()

Addressing unbalance

Can you think of any issues resulting from over/down sampling?

Final model

The “chosen” model

Let’s stick to the models without over/down sampling.

From the application exercise:

volcano_rec2 <- recipe(volcano_type ~ ., data = volcano_train) %>%
  step_other(tectonic_settings) %>%
  step_other(major_rock_1) %>%
  step_dummy(all_nominal_predictors()) %>%
  step_zv(all_predictors()) %>%
  step_center(all_predictors())

volcano_wflow2 <- workflow() %>%
  add_recipe(volcano_rec2) %>%
  add_model(volcano_spec)

Fitting the final model

final_fit <- last_fit(
  volcano_wflow2, 
  split = volcano_split
  )

collect_metrics(final_fit)
# A tibble: 2 × 4
  .metric  .estimator .estimate .config             
  <chr>    <chr>          <dbl> <chr>               
1 accuracy multiclass     0.629 Preprocessor1_Model1
2 roc_auc  hand_till      0.734 Preprocessor1_Model1

Confusion matrix

collect_predictions(final_fit) %>%
  conf_mat(volcano_type, .pred_class)
               Truth
Prediction      Stratovolcano Shield Other
  Stratovolcano            96     13    38
  Shield                    1      0     0
  Other                    21     16    55

Confusion matrix - visualized

collect_predictions(final_fit) %>%
  conf_mat(volcano_type, .pred_class) %>%
  autoplot()

ROC curve

collect_predictions(final_fit) %>%
  roc_curve(truth = volcano_type, .pred_Stratovolcano:.pred_Other) %>%
  autoplot()

ROC curve - altogether

Prediction

final_fitted <- extract_workflow(final_fit)

new_volcano <- tibble(
  latitude = 35.9940,
  longitude = -78.8986,
  elevation = 404,
  tectonic_settings = "Subduction zone / Continental crust (>25 km)",
  major_rock_1 = "Andesite / Basaltic Andesite"
)

predict(
  final_fitted, 
  new_volcano, 
  type = "prob"
  )
# A tibble: 3 × 1
  .pred_value
        <dbl>
1      0.381 
2      0.0379
3      0.581 

Acknowledgements

Inspired by

  • https://juliasilge.com/blog/multinomial-volcano-eruptions/
  • https://juliasilge.com/blog/nber-papers/