An introduction to the tidymodels package

# An introduction to the `tidymodels` package

## Bruna Wundervald, National University of Ireland, Maynooth

### Young-ISA Webinar
### January 28th, 2021

]

</div>

]

???

Welcome to the webinar on sharing on short notice

Where we'll show you how to get your teaching materials online with R Markdown.

---

---

name: clouds
class: center, middle
background-image: url(images/sea.jpg)
background-size: cover

.panelset .panel-tabs .panel-tab > a {
	color: #023d4d;
}
</style>

## .big-text[Hello]

### Bruna Wundervald

[GitHub: @brunaw](https://github.com/brunaw)  
[Twitter: @bwundervald](https://twitter.com/bwundervald)  
[Page: http://brunaw.com/](http://brunaw.com/)

---
class: middle, center

## .big-text[Today]

<div class="flex" style="margin: 0 1em;">
  <div class="column">
    <img src="https://raw.githubusercontent.com/rstudio/hex-stickers/master/SVG/tidymodels.svg"" style="width: 75%;">
  </div>
  
???

Here's who I know you are...

--
   <div class="column"style="margin: 0 1em;">
    <h3> Design matrices   </h3>
    <img src="https://raw.githubusercontent.com/rstudio/hex-stickers/master/SVG/recipes.svg" style="width: 65%;">
  </div>

???

--
   <div class="column"style="margin: 0 1em;">
    <h3> Resampling  </h3>
    <img src="https://raw.githubusercontent.com/rstudio/hex-stickers/master/SVG/rsample.svg" style="width: 65%;">
  </div>

???

--
   <div class="column"style="margin: 0 1em;">
    <h3> Model interfaces </h3>
    <img src="https://raw.githubusercontent.com/rstudio/hex-stickers/master/SVG/parsnip.svg" style="width: 65%;">
  </div>

???

--
   <div class="column"style="margin: 0 1em;">
    <h3> Tuning hyperparameters  </h3>
    <img src="https://raw.githubusercontent.com/rstudio/hex-stickers/master/SVG/tune.svg" style="width: 65%;">
  </div>

???

You have R Markdown files for teaching...

---
class: middle

# The `tidymodels` package

- Created by [Max Kuhn](https://github.com/topepo) and the `tidymodels` team

- Aims to be a unified collection of packages for modelling & machine
learning in `R`

- Easily integrates with the `tidyverse` packages

- Highly reusable infrastructure & reproducibility

> The packages presented today are only the **main** ones

---
background-color: #fef9c8

## Steps

1. Train and test separation with `rsample`
  2. Model specification and fitting with `parsnip`
  3. Feature engineering with `recipes`
  4. Hyperparameter tuning with `tune`
  
## Good practice: suffixes

- `_mod` for a `parsnip` model specification
- `_fit` for a fitted model
- `_rec` for a recipe
- `_tune` for a tuning object

---
class: middle 
.pull-left[  
## Data: Ames Housing

A data set from De Cock (2011) with 82 columns recorded for 2,930 properties in Ames IA.

Target variable: Sale price
Predictors: 
  - Location (e.g. neighborhood, lat and long) 
  - House elements (garage, year built, air conditioner,
 number of bedrooms/baths, etc)
]

]

---
class: middle

## Loading the data and the packages

```r
# Loading libraries
library(tidyverse)
library(tidymodels)

data(ames, package = "modeldata")
ames <- ames %>% mutate(Sale_Price = log10(Sale_Price))
```

---
class: middle

## A peek at the distribution of the sale prices

---
class: middle
background-color: #fef9c8

# 1. Train and test separation with `rsample`

```r
set.seed(2021)
data_split <- initial_split(ames, strata = "Sale_Price", prop = 0.8)

# Separating train and test
ames_train <- training(data_split)
ames_test  <- testing(data_split)
```
]

```r
data_split
## <Analysis/Assess/Total>
## <2346/584/2930>
```
]

```r
ames_train %>% slice(1:3)
## # A tibble: 3 x 74
##   MS_SubClass MS_Zoning Lot_Frontage Lot_Area Street Alley Lot_Shape
##   <fct>       <fct>            <dbl>    <int> <fct>  <fct> <fct>    
## 1 One_Story_… Resident…          141    31770 Pave   No_A… Slightly…
## 2 One_Story_… Resident…           80    11622 Pave   No_A… Regular  
## 3 One_Story_… Resident…           93    11160 Pave   No_A… Regular  
## # … with 67 more variables: Land_Contour <fct>, Utilities <fct>,
## #   Lot_Config <fct>, Land_Slope <fct>, Neighborhood <fct>, Condition_1 <fct>,
## #   Condition_2 <fct>, Bldg_Type <fct>, House_Style <fct>, Overall_Cond <fct>,
## #   Year_Built <int>, Year_Remod_Add <int>, Roof_Style <fct>, Roof_Matl <fct>,
## #   Exterior_1st <fct>, Exterior_2nd <fct>, Mas_Vnr_Type <fct>,
## #   Mas_Vnr_Area <dbl>, Exter_Cond <fct>, Foundation <fct>, Bsmt_Cond <fct>,
## #   Bsmt_Exposure <fct>, BsmtFin_Type_1 <fct>, BsmtFin_SF_1 <dbl>,
## #   BsmtFin_Type_2 <fct>, BsmtFin_SF_2 <dbl>, Bsmt_Unf_SF <dbl>,
## #   Total_Bsmt_SF <dbl>, Heating <fct>, Heating_QC <fct>, Central_Air <fct>,
## #   Electrical <fct>, First_Flr_SF <int>, Second_Flr_SF <int>,
## #   Gr_Liv_Area <int>, Bsmt_Full_Bath <dbl>, Bsmt_Half_Bath <dbl>,
## #   Full_Bath <int>, Half_Bath <int>, Bedroom_AbvGr <int>, Kitchen_AbvGr <int>,
## #   TotRms_AbvGrd <int>, Functional <fct>, Fireplaces <int>, Garage_Type <fct>,
## #   Garage_Finish <fct>, Garage_Cars <dbl>, Garage_Area <dbl>,
## #   Garage_Cond <fct>, Paved_Drive <fct>, Wood_Deck_SF <int>,
## #   Open_Porch_SF <int>, Enclosed_Porch <int>, Three_season_porch <int>,
## #   Screen_Porch <int>, Pool_Area <int>, Pool_QC <fct>, Fence <fct>,
## #   Misc_Feature <fct>, Misc_Val <int>, Mo_Sold <int>, Year_Sold <int>,
## #   Sale_Type <fct>, Sale_Condition <fct>, Sale_Price <dbl>, Longitude <dbl>,
## #   Latitude <dbl>
```
]
]

---
class: middle

# Resampling - more options:

- Bootstrap 
  - Cross-validation:
    - V-fold
    - Leave-one-out
    - Nested
    - Monte Carlo
---
class: middle
background-color: #fef9c8

# 2. Model specification and fitting with `parsnip`

1. Create a model specification: the type of model you want to run (lm, random forest, ...)
  2. Set an engine: the package used to run this model
  3. Fit the model
  
  
**List of available models:**  https://www.tidymodels.org/find/parsnip/  
---
class: middle

```r

model_setup <- rand_forest(mode = "regression", trees = 100)

rf_mod <- set_engine(model_setup, "ranger")

rf_fit <- fit(
  rf_mod, Sale_Price ~ Longitude + Latitude,
  data = ames_train
)
```
]

```
## Ranger result
## 
## Call:
##  ranger::ranger(formula = Sale_Price ~ Longitude + Latitude, data = data,      num.trees = ~100, regularization.factor = ~0.2, num.threads = 1,      verbose = FALSE, seed = sample.int(10^5, 1)) 
## 
## Type:                             Regression 
## Number of trees:                  100 
## Sample size:                      2346 
## Number of independent variables:  2 
## Mtry:                             1 
## Target node size:                 5 
## Variable importance mode:         none 
## Splitrule:                        variance 
## OOB prediction error (MSE):       0.009163911 
## R squared (OOB):                  0.7066803
```
]
]

---
class: middle

# `parsnip`: switching engines

```r
rf_mod <- model_setup %>% set_engine("randomForest")

rf_fit <- fit(
  rf_mod, Sale_Price ~ Longitude + Latitude,
  data = ames_train
)
```
]

```
## 
## Call:
##  randomForest(x = as.data.frame(x), y = y, ntree = ~100) 
##                Type of random forest: regression
##                      Number of trees: 100
## No. of variables tried at each split: 1
## 
##           Mean of squared residuals: 0.009115084
##                     % Var explained: 70.81
```
]
]

---
class: middle

```r
test_pred <- rf_fit %>% 
  predict(ames_test) %>% 
  bind_cols(ames_test)

rmse <- test_pred %>%  rmse(Sale_Price, .pred)
# (the rmse function comes from the yardstick package (!))
```
]

<img src="slides_files/figure-html/unnamed-chunk-13-1.png" width="504" style="display: block; margin: auto;" />
]
]

---
class: middle, center

---
class: middle
background-color: #fef9c8

# 3. Feature engineering with `recipes`

1. Create a `recipe()` to define the processing of the data, e.g.:
  - Create new classes, clean missing data, transform variables, etc
 2. Calculate that in the training set with the `prep()` function
 3. Apply the pre-processing with the `bake()` and 
 `juice()` functions 
  - The `bake()` function is used for 'new data', such as test sets
  
---
class: middle
.panelset[
.panel[.panel-name[The recipe]

```r
mod_rec <- recipe(
  Sale_Price ~ Longitude + Latitude + Neighborhood +
    Central_Air + Year_Built, 
  data = ames_train
) %>%
  # Factor levels that occur in <= 5% of data as "other"
  step_other(Neighborhood, threshold = 0.05) %>%
  # Create dummy variables for all factor variables
  step_dummy(all_nominal()) %>% 
  # Adds an interaction term
  step_interact(~ starts_with("Central_Air"):Year_Built) 
```
]

```
## Data Recipe
## 
## Inputs:
## 
##       role #variables
##    outcome          1
##  predictor          5
## 
## Operations:
## 
## Collapsing factor levels for Neighborhood
## Dummy variables from all_nominal()
## Interactions with starts_with("Central_Air"):Year_Built
```
]

```r
ames_rec <- prep(mod_rec, training = ames_train, verbose = TRUE)
## oper 1 step other [training] 
## oper 2 step dummy [training] 
## oper 3 step interact [training] 
## The retained training set is ~ 0.27 Mb  in memory.
```
]

```r
rf_mod <- rand_forest(
  mode = "regression", mtry = 5, trees = 500) %>% 
  set_engine("ranger", regularization.factor = 0.5)

rf_fit <- rf_mod %>% 
  fit(Sale_Price ~ ., data = juice(ames_rec)) 
```
]

```
## Ranger result
## 
## Call:
##  ranger::ranger(formula = Sale_Price ~ ., data = data, mtry = ~5,      num.trees = ~500, regularization.factor = ~0.5, num.threads = 1,      verbose = FALSE, seed = sample.int(10^5, 1)) 
## 
## Type:                             Regression 
## Number of trees:                  500 
## Sample size:                      2346 
## Number of independent variables:  12 
## Mtry:                             5 
## Target node size:                 5 
## Variable importance mode:         none 
## Splitrule:                        variance 
## OOB prediction error (MSE):       0.007954979 
## R squared (OOB):                  0.745376
```
]
]

---
class: middle, center

---
class: middle
background-color: #fef9c8

# 4. Hyperparameter tuning with `tune`

1. Choose the hyperparameters to tune
2. Choose the tuning method
3. Run and select the best hyperparameters

- We'll be doing grid search, but the package offers more options:
    - Bayesian optimization
    - Simulated annealing (`{finetune}` package)
    - Racing methods (`{finetune}` package)

---
class: middle

```r
ctrl <- control_grid(save_pred = TRUE)

rf_mod <- rand_forest(mtry = tune()) %>%
  set_mode("regression") %>%
  set_engine("ranger", regularization.factor = tune())

rf_param <- parameters(rf_mod)
```
]

```
## Collection of 2 parameters for tuning
## 
##                     id        parameter type object class
##                   mtry                  mtry    nparam[?]
##  regularization.factor regularization.factor    nparam[+]
## 
## Model parameters needing finalization:
##    # Randomly Selected Predictors ('mtry')
## 
## See `?dials::finalize` or `?dials::update.parameters` for more information.
```
]

```r
set.seed(2021)
data_folds <- vfold_cv(data = juice(ames_rec), v = 5)
ranger_tune <-
  rf_mod %>%
  tune_grid(
    Sale_Price ~ ., 
    resamples = data_folds,
    grid = 10,
    control = ctrl
  )
```
]
]

---
class: middle

## Evaluating performance results

<img src="slides_files/figure-html/unnamed-chunk-24-1.png" width="40%" height="30%" style="display: block; margin: auto;" />
]

<img src="slides_files/figure-html/unnamed-chunk-25-1.png" width="40%" height="30%" style="display: block; margin: auto;" />
]

]

---
class: middle

## Using the best hyperparameters

```r
best_res <- select_best(ranger_tune, metric = "rmse")
final_rf_mod <- 
  rand_forest(mtry = best_res$mtry) %>%
  set_mode("regression") %>%
  set_engine("ranger", 
             regularization.factor = best_res$regularization.factor)

final_rf_fit <- final_rf_mod %>%
  fit(Sale_Price ~ ., data = juice(ames_rec))
```
]

```r
test_bake <- bake(ames_rec, new_data = ames_test)
final_pred <- final_rf_fit %>% 
  predict(test_bake) %>% 
  bind_cols(ames_test)

final_rmse <- final_pred %>%  rmse(Sale_Price, .pred)
```
]

<img src="slides_files/figure-html/unnamed-chunk-28-1.png" width="40%" height="30%" style="display: block; margin: auto;" />
]
]

---

## Resources

- [Talk GitHub repository](https://github.com/brunaw/tidymodels-webinar)
- http://tidymodels.org/
  - Tutorials and function documentation
- [Book: Tidy Modeling with `R`](https://www.tmwr.org/)
- [Book: Applied Predictive Modeling](http://appliedpredictivemodeling.com/)
-  [Code: Applied Predictive Modeling](https://github.com/topepo/tidyAPM)

- Max's workshops:
  - https://github.com/topepo/2020-earl-workshop
  - https://github.com/topepo/RPharma-2019-Workshop
  - https://github.com/topepo/nyr-2020
  
---
background-image: url(images/sea.jpg)
background-size: cover
class: center, middle, inverse

## .big-text[Questions?]

---

## Thank you!

### Find me at...
[GitHub: @brunaw](https://github.com/brunaw)  
[Twitter: @bwundervald](https://twitter.com/bwundervald)

[Page: http://brunaw.com/](http://brunaw.com/)

Slides template by [Dr. Alison Hill](http://twitter.com/apreshill) & 
illustrations by [Allison Horst](https://github.com/allisonhorst/stats-illustrations)