LASSO Models

Published

2025-10-24

We’ve seen that using a training/test split can give us a way to judge our model’s effectiveness on data it wasn’t trained on.

Alternatively, we’ve seen that cross-validation can be used to judge our model’s effectiveness without doing the training/test split. We might do this if we have a smaller data set.

A third option is to use both CV and a training/test split. We might do this if we had a number of different classes or families of models we were fitting. We could use CV on the training data only to choose the best from each family. We refit each of those ‘best’ models on the full training data. Then we could compare those ‘best’ models from each class on the test set to select our final best model! (We would then refit that model to the entire data set.)

For instance, we might have the following families of models:

MLR models (say some with interactions, some without)
A LASSO model (type of MLR model with penalization)
A random forest model (studied later)
A kNN (k nearest neighbor) model (we won’t study this one)

The latter three model types all have tuning parameters that must be selected when fitting the model. CV on the training set is often used to select those tuning parameters! We can also use CV on the training set to determine the best MLR model (or use some other model selection technique).

Using this process, we would get a best model of each type (or family) using the training set only. Then we refit those models on the full training data and can compare the best model from each family of models on the test set to select an overall best model!

Let’s introduce the LASSO model and then use CV (on the training set) to select the tuning parameter. We’ll similarly compare some competing MLR models using CV on the training set. Once we have our best LASSO and best MLR model, we’ll then go to the test set to determine an overall best model.

Fitting Models on our `bike_data`

First, a training/test split on our bike_data:

library(tidyverse)
library(tidymodels)
set.seed(10)
bike_data <- read_csv("https://www4.stat.ncsu.edu/~online/datasets/bikeDetails.csv")
bike_data <- bike_data |> 
  mutate(log_selling_price = log(selling_price), 
         log_km_driven = log(km_driven),
         owners = ifelse(owner == "1st owner", "single", "multiple")) |>
  select(log_km_driven, log_selling_price, everything())
#use tidymodel functions for splitting the data
bike_split <- initial_split(bike_data, prop = 0.7)
bike_train <- training(bike_split)
bike_test <- testing(bike_split)

Now we’ll proceed with finding our best MLR model and best LASSO model on just the bike_train data.

Selecting the Best MLR Model Using CV

Again, it isn’t necessary to use CV to select an MLR model (there are tons of ways to choose between candidate MLR models - AIC, BIC, backward selection, Mallow’s Cp, …). However, CV is a perfectly valid way to choose between MLR models.

Let’s compare the following models

MLR model 1: log_selling_price ~ log_km_driven + owners + year (fits different intercept based on owners being “single” or “multiple” and slope terms for log_km_driven and year)
MLR model 2: log_selling_price ~ log_km_driven + owners + log_km_driven*owners (fits different SLR models between log_selling_price and log_km_driven (different slopes and intercepts) for each setting of owners)
MLR model 3: log_selling_price ~ (log_km_driven + owners + year)^2 (fits different a model with all pairwise interactions and main effects - essentially two separate MLR models for each setting of owners using log_km_driven and year as predictors)

We want to just use the training data here and 10 fold CV. First, let’s create our CV folds:

#create folds
bike_CV_folds <- vfold_cv(bike_train, 10)

Remember to use tidymodels we want to set up our model (and engine) and create our recipes. Those go into a workflow that can then be fit on the above folds.

Let’s define our basic linear model:

#set up how we'll fit our linear model
MLR_spec <- linear_reg() |>
  set_engine("lm")

Now define our recipes for the three models. Note: we created the log_km_driven variable earlier. This is ok because this isn’t a learned transform. That is, there is no danger of training/test set data leakage issues since we know how to do the natural log transform, regardless of the data.

#define our MLR models
MLR_recipe1 <- recipe(log_selling_price ~ log_km_driven + owners + year, 
                      data = bike_train) |>
  step_dummy(owners)
MLR_recipe2 <- recipe(log_selling_price ~ log_km_driven + owners,
                      data = bike_train) |>
  step_dummy(owners) |>
  step_interact(~log_km_driven:starts_with("owner"))

MLR_recipe3 <- recipe(log_selling_price ~ log_km_driven + owners + year,
                      data = bike_train) |>
  step_dummy(owners) |>
  step_interact(~log_km_driven:starts_with("owner") + log_km_driven:year + starts_with("owner"):year)

Now we create our workflows for each model:

MLR_wkf1 <- workflow() |>
  add_recipe(MLR_recipe1) |>
  add_model(MLR_spec)

MLR_wkf2 <- workflow() |>
  add_recipe(MLR_recipe2) |>
  add_model(MLR_spec)

MLR_wkf3 <- workflow() |>
  add_recipe(MLR_recipe3) |>
  add_model(MLR_spec)

Let’s fit these models to our CV folds and see how they perform!

MLR_fit1 <-  MLR_wkf1 |>
  fit_resamples(bike_CV_folds)

MLR_fit2 <- MLR_wkf2 |>
  fit_resamples(bike_CV_folds) 

MLR_fit3 <- MLR_wkf3 |>
  fit_resamples(bike_CV_folds)

Combine the metrics across the folds and create a final data frame with the results

rbind(MLR_fit1 |> collect_metrics() |> filter(.metric == "rmse"),
      MLR_fit2 |> collect_metrics() |> filter(.metric == "rmse"),
      MLR_fit3 |> collect_metrics() |> filter(.metric == "rmse")) |> 
  mutate(Model = c("Model 1", "Model 2", "Model 3")) |>
  select(Model, mean, n, std_err)

# A tibble: 3 × 4
  Model    mean     n std_err
  <chr>   <dbl> <int>   <dbl>
1 Model 1 0.494    10  0.0200
2 Model 2 0.570    10  0.0158
3 Model 3 0.484    10  0.0180

Based on RMSE, we see that the last model is the best MLR model of the three we fit! Note again, we’ve chosen between the three models from the family of models (MLR models) using just CV on the training data.

Let’s refit that on the entire training set.

MLR_final <-  MLR_wkf3 |>
  fit(bike_train)
tidy(MLR_final)

# A tibble: 7 × 5
  term                            estimate std.error statistic       p.value
  <chr>                              <dbl>     <dbl>     <dbl>         <dbl>
1 (Intercept)                    -74.2      87.0        -0.853 0.394        
2 log_km_driven                    3.42      8.58        0.398 0.691        
3 year                             0.0442    0.0433      1.02  0.308        
4 owners_single                 -136.       22.4        -6.09  0.00000000183
5 log_km_driven_x_owners_single    0.222     0.0696      3.19  0.00148      
6 log_km_driven_x_year            -0.00190   0.00427    -0.445 0.657        
7 owners_single_x_year             0.0665    0.0110      6.02  0.00000000273

Fit a LASSO Model Using CV

A similar model to the MLR model that involves a tuning parameter is the LASSO model.

Least Angle Subset and Selection Operator or LASSO
- Similar to Least Squares but a penalty is placed on the sum of the absolute values of the regression coefficients
- $\alpha$ (>0) is called a tuning parameter
$\min\limits_{\beta's}\sum_{i=1}^{n}(y_i-(\beta_0+\beta_1x_{1i}+...+\beta_px_{pi}))^2 + \alpha\sum_{j=1}^{p}|\beta_j|$
- For an $\alpha$ of 0 we get the usual least squares estimates. As $\alpha$ gets larger we ‘shrink’ the estimates toward 0.

We have a different model fit for each value of $\alpha$ . When choosing the tuning parameter ( $\alpha$ ), we are really considering a family of models for a given set of predictors!
- Consider the MLR model with different intercepts based on owners that includes both log_km_driven and year as predictors. The plot here gives the coefficient profiles (estimates) as a function of $\alpha$ (although and $\alpha$ of 0 here corresponds to being on the right hand side of the plot).

To read this plot, note that each line represents the value of one of the ‘slope’ coefficients in the model. All the way on the right we have the unconstrained solution (the usual MLR solution). Compare the values there to

lm(log_selling_price ~ log_km_driven + year + owners, data = bike_train)$coef

  (Intercept) log_km_driven          year  ownerssingle 
-134.57077580   -0.24380816    0.07339603   -0.08397016

As we move to the left on the graph (as $\alpha$ increases) we see that our coefficients are shrunk towards 0, eventually being set to 0!

We want to choose which level of shrinkage is appropriate. That is, which value of $\alpha$ gives us the best model for predicting! A perfect case for using CV!

Create our Recipe

Let’s use 10 fold CV to choose our $\alpha$ value (i.e. figure out our best LASSO model with these three predictors). As we are penalizing the magnitude of the slope estimates (and the estimates depend on the units of the predictor), LASSO models should really be fit on standardized predictors. Let’s include that step in our recipe!

#set up how we'll fit our LASSO model
#code modified from https://juliasilge.com/blog/lasso-the-office/
LASSO_recipe <- recipe(log_selling_price ~ log_km_driven + owners + year, 
                      data = bike_train) |>
  step_dummy(owners) |>
  step_normalize(log_km_driven, year)

Create a Model Instance with `tune()`

Great, now we need to specify a LASSO model. Recall this page allows us to find a type of model to fit along with how to specify it and the engine used to fit the model.

In this case, we still want linear_reg() but we want to use the ‘glmnet’ engine. We also want to specify the penalty parameter (corresponds to a form of $\alpha$ we mentioned above). “glmnet” actually allows us to fit a more complicated model (the elastic net) so there is a second tuning parameter to deal with (called mixture).

We set mixture = 1 to turn this into a LASSO model (rather than an elastic net model)
We set penalty = tune() to tell tidymodels we are going to use a resampling method to choose this parameter

LASSO_spec <- linear_reg(penalty = tune(), mixture = 1) |>
  set_engine("glmnet")

Sweet - now we create our workflow.

LASSO_wkf <- workflow() |>
  add_recipe(LASSO_recipe) |>
  add_model(LASSO_spec)
LASSO_wkf

══ Workflow ════════════════════════════════════════════════════════════════════
Preprocessor: Recipe
Model: linear_reg()

── Preprocessor ────────────────────────────────────────────────────────────────
2 Recipe Steps

• step_dummy()
• step_normalize()

── Model ───────────────────────────────────────────────────────────────────────
Linear Regression Model Specification (regression)

Main Arguments:
  penalty = tune()
  mixture = 1

Computational engine: glmnet

Fit the Model with `tune_grid()` and `grid_regular()`

We saw how to fit a workflow to a set of CV folds with fit_resample(). Since we have a tuning parameter here, we don’t want to use that function. Instead, we use tune_grid(). This function allows us to fit the model to CV folds but specify the set of tuning parameters to consider.

This implies we are actually doing a bunch of model fits on the CV folds! One for each tuning parameter we specify.
In the tune_grid() function we can specify the values of the tuning parameter with the grid = argument.
grid_regular() is a function that can be used to choose a grid of reasonable values

#A warning will occur for one value of the tuning parameter, safe to ignore
LASSO_grid <- LASSO_wkf |>
  tune_grid(resamples = bike_CV_folds,
            grid = grid_regular(penalty(), levels = 200)) 
LASSO_grid

# Tuning results
# 10-fold cross-validation 
# A tibble: 10 × 4
   splits           id     .metrics           .notes          
   <list>           <chr>  <list>             <list>          
 1 <split [667/75]> Fold01 <tibble [400 × 5]> <tibble [8 × 4]>
 2 <split [667/75]> Fold02 <tibble [400 × 5]> <tibble [8 × 4]>
 3 <split [668/74]> Fold03 <tibble [400 × 5]> <tibble [8 × 4]>
 4 <split [668/74]> Fold04 <tibble [400 × 5]> <tibble [8 × 4]>
 5 <split [668/74]> Fold05 <tibble [400 × 5]> <tibble [8 × 4]>
 6 <split [668/74]> Fold06 <tibble [400 × 5]> <tibble [8 × 4]>
 7 <split [668/74]> Fold07 <tibble [400 × 5]> <tibble [8 × 4]>
 8 <split [668/74]> Fold08 <tibble [400 × 5]> <tibble [8 × 4]>
 9 <split [668/74]> Fold09 <tibble [400 × 5]> <tibble [8 × 4]>
10 <split [668/74]> Fold10 <tibble [400 × 5]> <tibble [8 × 4]>

There were issues with some computations:

  - Warning(s) x80: A correlation computation is required, but `estimate` is constant...

Run `show_notes(.Last.tune.result)` for more information.

This may look like it only has one fit per fold but the list columns hide the fact that we actually have fit 200 separate LASSO models (one for each tuning parameter specified by grid_regular()). Notice this:

LASSO_grid[1, ".metrics"][[1]]

[[1]]
# A tibble: 400 × 5
    penalty .metric .estimator .estimate .config          
      <dbl> <chr>   <chr>          <dbl> <chr>            
 1 1   e-10 rmse    standard       0.475 pre0_mod001_post0
 2 1   e-10 rsq     standard       0.596 pre0_mod001_post0
 3 1.12e-10 rmse    standard       0.475 pre0_mod002_post0
 4 1.12e-10 rsq     standard       0.596 pre0_mod002_post0
 5 1.26e-10 rmse    standard       0.475 pre0_mod003_post0
 6 1.26e-10 rsq     standard       0.596 pre0_mod003_post0
 7 1.41e-10 rmse    standard       0.475 pre0_mod004_post0
 8 1.41e-10 rsq     standard       0.596 pre0_mod004_post0
 9 1.59e-10 rmse    standard       0.475 pre0_mod005_post0
10 1.59e-10 rsq     standard       0.596 pre0_mod005_post0
# ℹ 390 more rows

This is actually a tibble with 400 metrics (200 rmse values and 200 rsq values).

Of course, we want to have these metrics computed across the folds for each of our 200 values of the tuning parameter. We use collect_metrics() as before!

LASSO_grid |>
  collect_metrics() |>
  filter(.metric == "rmse")

# A tibble: 200 × 7
    penalty .metric .estimator  mean     n std_err .config          
      <dbl> <chr>   <chr>      <dbl> <int>   <dbl> <chr>            
 1 1   e-10 rmse    standard   0.494    10  0.0199 pre0_mod001_post0
 2 1.12e-10 rmse    standard   0.494    10  0.0199 pre0_mod002_post0
 3 1.26e-10 rmse    standard   0.494    10  0.0199 pre0_mod003_post0
 4 1.41e-10 rmse    standard   0.494    10  0.0199 pre0_mod004_post0
 5 1.59e-10 rmse    standard   0.494    10  0.0199 pre0_mod005_post0
 6 1.78e-10 rmse    standard   0.494    10  0.0199 pre0_mod006_post0
 7 2.00e-10 rmse    standard   0.494    10  0.0199 pre0_mod007_post0
 8 2.25e-10 rmse    standard   0.494    10  0.0199 pre0_mod008_post0
 9 2.52e-10 rmse    standard   0.494    10  0.0199 pre0_mod009_post0
10 2.83e-10 rmse    standard   0.494    10  0.0199 pre0_mod010_post0
# ℹ 190 more rows

Ok, but it is tough to see the values there. Let’s plot it instead.

LASSO_grid |>
  collect_metrics() |>
  filter(.metric == "rmse") |>
  ggplot(aes(penalty, mean, color = .metric)) +
  geom_line()

For the smallest value of the penalty (left side of the graph) we have no shrinkage. As our penalty gets bigger we have to shrink our coefficients more (further right on the graph has more shrinkage). Having little to no shrinkage is associated with the lower RMSE values based on our CV results!

Pull Out the ‘Best’ Model with `select_best()` and `finalize_workflow()`

We can get the tuning parameter corresponding to the best RMSE value and determine which coefficients that model has using select_best() and finalize_workflow().

lowest_rmse <- LASSO_grid |>
  select_best(metric = "rmse")
lowest_rmse

# A tibble: 1 × 2
  penalty .config          
    <dbl> <chr>            
1  0.0174 pre0_mod165_post0

Now fit that ‘best’ LASSO model on the entire training set. finalize_workflow() tells R to finish our training with a specific setting of the terms we set to tune() in our model definition. We can supply the result from the previous code chunk to get the best model.

LASSO_wkf |>
  finalize_workflow(lowest_rmse)

══ Workflow ════════════════════════════════════════════════════════════════════
Preprocessor: Recipe
Model: linear_reg()

── Preprocessor ────────────────────────────────────────────────────────────────
2 Recipe Steps

• step_dummy()
• step_normalize()

── Model ───────────────────────────────────────────────────────────────────────
Linear Regression Model Specification (regression)

Main Arguments:
  penalty = 0.0174263338600965
  mixture = 1

Computational engine: glmnet

#fit it to the entire training set to see the model fit
LASSO_final <- LASSO_wkf |>
  finalize_workflow(lowest_rmse) |>
  fit(bike_train)
tidy(LASSO_final)

# A tibble: 4 × 3
  term          estimate penalty
  <chr>            <dbl>   <dbl>
1 (Intercept)    10.7     0.0174
2 log_km_driven  -0.233   0.0174
3 year            0.307   0.0174
4 owners_single  -0.0151  0.0174

Comparing Chosen Models on the Test Set

Now we have our best model from each ‘family’ of models (one MLR and one LASSO model). Let’s take them to the test set and see how they perform on this data!

We can use last_fit() on the bike_split object as we did in the previous section of notes
This uses the training set transformations on the test set and does predictions

MLR_wkf3 |>
  last_fit(bike_split) |>
  collect_metrics()

# A tibble: 2 × 4
  .metric .estimator .estimate .config        
  <chr>   <chr>          <dbl> <chr>          
1 rmse    standard       0.546 pre0_mod0_post0
2 rsq     standard       0.483 pre0_mod0_post0

LASSO_wkf |>
  finalize_workflow(lowest_rmse) |>
  last_fit(bike_split) |>
  collect_metrics()

# A tibble: 2 × 4
  .metric .estimator .estimate .config        
  <chr>   <chr>          <dbl> <chr>          
1 rmse    standard       0.557 pre0_mod0_post0
2 rsq     standard       0.473 pre0_mod0_post0

We see that the MLR model outperforms the LASSO model! This would be our overall best model.

Just to relate this to how we found this previously, note that we can do this ourselves in the usual manner. That is, we can use our training set fit (MLR_final and LASSO_final) with predict() and rmse_vec() from yardstick:

MLR_final |>
  predict(bike_test) |>
  pull() |>
  rmse_vec(truth = bike_test$log_selling_price)

[1] 0.5460833

LASSO_final |>
  predict(bike_test) |>
  pull() |>
  rmse_vec(truth = bike_test$log_selling_price)

[1] 0.5565673

As MLR_final and LASSO_final both have class workflow, using predict() actually uses predict.workflow(). This means it does the appropriate training set transformations prior to predicting on the test set! From the documentation for predict.workflow():

This is the predict() method for a fit workflow object. The nice thing about predicting from a workflow is that it will:

Preprocess new_data using the preprocessing method specified when the workflow was created and fit. This is accomplished using hardhat::forge(), which will apply any formula preprocessing or call recipes::bake() if a recipe was supplied.

Call parsnip::predict.model_fit() for you using the underlying fit parsnip model.

Final Step

Now that we have an overall best model from our set of best models :) We would now refit the best model on the full dataset for future use. Again, we want to apply the transformations laid out previously. However, any transforms that depend on the data should now be based on the full data, not just the training data. By using tidymodels this is taken care of for us! We simply fit() the model with the full data set.

final_model <- MLR_wkf3 |>
  fit(bike_data) 
tidy(final_model)

# A tibble: 7 × 5
  term                            estimate std.error statistic       p.value
  <chr>                              <dbl>     <dbl>     <dbl>         <dbl>
1 (Intercept)                   -127.       79.6        -1.60  0.110        
2 log_km_driven                    6.56      7.82        0.839 0.402        
3 year                             0.0709    0.0396      1.79  0.0738       
4 owners_single                 -126.       20.8        -6.07  0.00000000179
5 log_km_driven_x_owners_single    0.281     0.0616      4.57  0.00000555   
6 log_km_driven_x_year            -0.00348   0.00389    -0.895 0.371        
7 owners_single_x_year             0.0613    0.0103      5.97  0.00000000333

If we want the final model fit in the usual lm form, we can use extract_fit_parsnip()

almost_usual_fit <- extract_fit_parsnip(final_model)
usual_fit <- almost_usual_fit$fit
summary(usual_fit)


Call:
stats::lm(formula = ..y ~ ., data = data)

Residuals:
     Min       1Q   Median       3Q      Max 
-1.82178 -0.34620 -0.05771  0.26085  2.50456 

Coefficients:
                                Estimate Std. Error t value Pr(>|t|)    
(Intercept)                   -1.274e+02  7.964e+01  -1.600   0.1099    
log_km_driven                  6.557e+00  7.817e+00   0.839   0.4018    
year                           7.089e-02  3.961e-02   1.790   0.0738 .  
owners_single                 -1.263e+02  2.081e+01  -6.069 1.79e-09 ***
log_km_driven_x_owners_single  2.811e-01  6.157e-02   4.566 5.55e-06 ***
log_km_driven_x_year          -3.482e-03  3.889e-03  -0.895   0.3708    
owners_single_x_year           6.127e-02  1.027e-02   5.965 3.33e-09 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 0.5001 on 1054 degrees of freedom
Multiple R-squared:  0.5081,    Adjusted R-squared:  0.5053 
F-statistic: 181.4 on 6 and 1054 DF,  p-value: < 2.2e-16

Wrap-up

If you are only considering one type of model, you can use just a training/test set or just use k-fold CV to select the best version of that model
When you have multiple types of models to choose from, we usually use both! Why?
- When we use the test set too much, we may have ‘data leakage’
- Essentially we end up training our models to the test set by using it too much
- Using CV on the training set and using the test set as a final comparison helps us avoid this!

Recap

Cross-validation gives a way to use more of the data while still seeing how the model does on test data

Commonly 5 fold or 10 fold is done
Once a best model is chosen, model is refit on entire data set

We can use CV with or without a training/test split, depending on how much data we have and whether or not we have tuning parameters!

Use the table of contents on the left or the arrows at the bottom of this page to navigate to the next learning material!

Fitting Models on our bike_data