STA 9750 – STA 9750

Agenda

Predictive Modeling with tidymodels

Adapted from (Case Study)[https://www.tidymodels.org/start/case-study/]

Breakout Rooms

Order	Team	Order	Team
1	Rat Pack	6	Ca$h VZ
2	Subway Surfers	7	Listing Legends
3	Chart Toppers	8	TDSSG
4	Metro Mindset	9	Broker T’s
5	Apple Watch	10	EVengers

`tidymodels`

Strength of R:

Thousands of authors contributing packages to CRAN

Weakness of R:

Thousands of authors contributing slightly incompatible packages to CRAN

No two modeling packages have exactly the same API. Makes changing between interfaces cumbersome

`tidymodels`

tidymodels attemps to provide a uniform interface to a wide variety of predictive Machine Learning tools

Advantages:

Easy to swap out different algorithms to find the best

Disadvantages:

Harder to take advantage of the strengths of each approach

I have dedicated my academic life to the differences in these methods, but 99% of the time, “black-box” prediction is good enough. In STA 9890, we get into the weeds - not here.

ML vs Statistical Pipelines

Statistics / Data Science:

Find the model that fits the data best
Model should capture all important data features
Interpretability
History: Grounded in lab sciences where experiments are expensive and data is limited

ML vs Statistical Pipelines

Machine Learning:

Find the model that predicts the data best
No “perfect” model - just the best one we’ve found so far
Black-box techniques are great, if effective
History: Silicon Valley “at scale”

Validation based on of-of-sample or test predictions

Validating Predictive Power

How to check whether a model predicts well?

Need more data! But where to get more data?

Actually get more data (hard, expensive, slow)
Split data into parts - test/training split
Cross-Validation
Resampling

Today, we’ll primarily use a combination: Test/Train split & Cross-Validation!

Cross-Validation

Cross-Validation is done on the estimator, not the fitted algorithm

`tidymodels`

tidymodels workflow:

Initial Split
Pre-Process
Fit (many) models
Select best
Refit
Test Set Assessment

tidymodels is very punny, so a bit hard to tell which step is which…

Initial Split

# Stratified sampling to ensure balance
splits      <- initial_split(hotels, 
                             strata = children)

hotel_train <- training(splits)
hotel_test  <- testing(splits)

Pre-Process

holidays <- c("AllSouls", "AshWednesday", "ChristmasEve", "Easter", 
              "ChristmasDay", "GoodFriday", "NewYearsDay", "PalmSunday")

recipe <- 
  recipe(children ~ ., data = hotel_other) |> 
  step_date(arrival_date) |> 
  step_holiday(arrival_date, holidays = holidays) |> 
  step_rm(arrival_date) |> 
  step_dummy(all_nominal_predictors()) |> 
  step_zv(all_predictors()) |> 
  step_normalize(all_predictors())

Fit Models

lr_model <- 
  logistic_reg(penalty = tune(), mixture = 1) |> 
  set_engine("glmnet")

Select Best

Find a grid of parameters

lr_reg_grid <- data.frame(penalty = 10^seq(-4, -1, length.out = 30))

Perform CV splits:

lr_folds <- vfold_cv(hotel_train, v = 5)

Select Best

Define a workflow:

lr_workflow <-  
  workflow() |> 
  add_model(lr_mod) |> 
  add_recipe(lr_recipe)

Fit workflow to a grid of parameters:

lr_results <- 
  lr_workflow |> 
  tune_grid(grid = lr_reg_grid,
            control = control_grid(save_pred = TRUE),
            resamples = lr_folds,
            metrics = metric_set(roc_auc))

Select Best

Visual examination

lr_results |> 
  collect_metrics() |> 
  ggplot(aes(x = penalty, y = mean)) + 
  geom_point() + 
  geom_line() + 
  ylab("Area under the ROC Curve") +
  scale_x_log10(labels = scales::label_number())

Select Best

lr_results |> show_best()
lr_best <- lr_results |> select_best()

lr_best_fit <- lr_results |> fit_best()

Refit

lr_best_fit <- lr_results |> fit_best()

Test Set Assessment

predict(lr_best_fit, hotel_test)

Exercise

Work through the random forest components of https://www.tidymodels.org/start/case-study

You’ll need to work through the data import elements as well

Other `tidymodels` tools

Model Stacking
Probabilistic Predictions
Uncertainty Bounds (Conformal Inference)
Multilevel (Mixed-Effect) Models
Fairness Audits

STA 9750 - Week 13

STA 9750 Mini-Project #04

Week 13 Pre-Assignment

Week 13 Pre-Assignment

Grading

Upcoming

Final Project Presentations

Final Project Presentations

Final Project Reports

Peer Assessment

Final Project Grading

Today

Agenda

Breakout Rooms

`tidymodels`

`tidymodels`

ML vs Statistical Pipelines

ML vs Statistical Pipelines

Validating Predictive Power

Cross-Validation

`tidymodels`

Initial Split

Pre-Process

Fit Models

Select Best

Select Best

Select Best

Select Best

Refit

Test Set Assessment

Exercise

Other `tidymodels` tools

More Reading