Statistics, Science, Random Ramblings

A blog about data and other interesting things

An introduction to tidymodels

Posted at — Mar 3, 2020

Recently, the tidymodels family of packages has gained some attention, although some of the included packages have been around for a while. However, as of February 2020 many of the involved packages still have a version number of 0.0.x, so it is likely that things will change quite a bit in the future.

What is tidymodels?

What the tidyverse family of packages is for data wrangling, the tidymodels family aims to be for modelling. Additionally, what scikit-learn is for Python, tidymodels could be for R. There are numerous packages available that implement various kinds of modelling and predicting data available in R, all with their own properties, quirks and features. This is where tidymodels comes in, as it provides a standardised interface to various modelling packages, so modelling and predicting becomes a straightforward process.

In this blog post I will go through several steps of using tidymodels from data preprocessing to hyperparameter tuning. The data used is a dataset on bike rentals in Washington D.C., USA in 2011 and 2012.

If you want to skip the data exploration and data wrangling part below, click here to jump directly to the preprocessing section.

Data exploration

Load the required packages and set a seed for reproducibility. Similar as with tidyverse, calling library("tidymodels") will load many packages at once.

library("tidyverse")
library("tidymodels")
set.seed(4757)

Of course, first we should read the data set.

bikes <- read_csv("bikes_hour.csv")

And now have a look at it.

glimpse(bikes)
## Observations: 17,379
## Variables: 17
## $ instant    <dbl> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,…
## $ dteday     <date> 2011-01-01, 2011-01-01, 2011-01-01, 2011-01-01, 2011-01-0…
## $ season     <dbl> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1…
## $ yr         <dbl> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…
## $ mnth       <dbl> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1…
## $ hr         <dbl> 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, …
## $ holiday    <dbl> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…
## $ weekday    <dbl> 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6…
## $ workingday <dbl> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…
## $ weathersit <dbl> 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 3, 3…
## $ temp       <dbl> 0.24, 0.22, 0.22, 0.24, 0.24, 0.24, 0.22, 0.20, 0.24, 0.32…
## $ atemp      <dbl> 0.2879, 0.2727, 0.2727, 0.2879, 0.2879, 0.2576, 0.2727, 0.…
## $ hum        <dbl> 0.81, 0.80, 0.80, 0.75, 0.75, 0.75, 0.80, 0.86, 0.75, 0.76…
## $ windspeed  <dbl> 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0896, 0.0000, 0.…
## $ casual     <dbl> 3, 8, 5, 3, 0, 0, 2, 1, 1, 8, 12, 26, 29, 47, 35, 40, 41, …
## $ registered <dbl> 13, 32, 27, 10, 1, 1, 0, 2, 7, 6, 24, 30, 55, 47, 71, 70, …
## $ cnt        <dbl> 16, 40, 32, 13, 1, 1, 2, 3, 8, 14, 36, 56, 84, 94, 106, 11…

There are the following columns:

  • instant: a unique id for each record
  • dteday: the date of the record
  • season: 1 to 4, spring to winter
  • yr, mnth, hr: year, month, hour
  • holiday, weekday, workingday: whether it is a holiday, weekday or working day
  • weathersit: 1 to 4, with 1 being clear weather and 4 heavy precipitation
  • temp, atemp: temperature and feels-like temperature, divided by the maximum for each variable
  • hum: relative humidity divided by 100 (the maximum)
  • windspeed: windspeed divided by 67 (the maximum), unit is unclear but sounds like mph
  • casual, registered, cnt: amount of rentals by casual/registered/total users

Are there missings?

sum(is.na(bikes))
## [1] 0

No missings, that’s always nice. Let’s further look whether something in the data seems off:

summary(bikes)
##     instant          dteday               season            yr        
##  Min.   :    1   Min.   :2011-01-01   Min.   :1.000   Min.   :0.0000  
##  1st Qu.: 4346   1st Qu.:2011-07-04   1st Qu.:2.000   1st Qu.:0.0000  
##  Median : 8690   Median :2012-01-02   Median :3.000   Median :1.0000  
##  Mean   : 8690   Mean   :2012-01-02   Mean   :2.502   Mean   :0.5026  
##  3rd Qu.:13034   3rd Qu.:2012-07-02   3rd Qu.:3.000   3rd Qu.:1.0000  
##  Max.   :17379   Max.   :2012-12-31   Max.   :4.000   Max.   :1.0000  
##       mnth              hr           holiday           weekday     
##  Min.   : 1.000   Min.   : 0.00   Min.   :0.00000   Min.   :0.000  
##  1st Qu.: 4.000   1st Qu.: 6.00   1st Qu.:0.00000   1st Qu.:1.000  
##  Median : 7.000   Median :12.00   Median :0.00000   Median :3.000  
##  Mean   : 6.538   Mean   :11.55   Mean   :0.02877   Mean   :3.004  
##  3rd Qu.:10.000   3rd Qu.:18.00   3rd Qu.:0.00000   3rd Qu.:5.000  
##  Max.   :12.000   Max.   :23.00   Max.   :1.00000   Max.   :6.000  
##    workingday       weathersit         temp           atemp       
##  Min.   :0.0000   Min.   :1.000   Min.   :0.020   Min.   :0.0000  
##  1st Qu.:0.0000   1st Qu.:1.000   1st Qu.:0.340   1st Qu.:0.3333  
##  Median :1.0000   Median :1.000   Median :0.500   Median :0.4848  
##  Mean   :0.6827   Mean   :1.425   Mean   :0.497   Mean   :0.4758  
##  3rd Qu.:1.0000   3rd Qu.:2.000   3rd Qu.:0.660   3rd Qu.:0.6212  
##  Max.   :1.0000   Max.   :4.000   Max.   :1.000   Max.   :1.0000  
##       hum           windspeed          casual         registered   
##  Min.   :0.0000   Min.   :0.0000   Min.   :  0.00   Min.   :  0.0  
##  1st Qu.:0.4800   1st Qu.:0.1045   1st Qu.:  4.00   1st Qu.: 34.0  
##  Median :0.6300   Median :0.1940   Median : 17.00   Median :115.0  
##  Mean   :0.6272   Mean   :0.1901   Mean   : 35.68   Mean   :153.8  
##  3rd Qu.:0.7800   3rd Qu.:0.2537   3rd Qu.: 48.00   3rd Qu.:220.0  
##  Max.   :1.0000   Max.   :0.8507   Max.   :367.00   Max.   :886.0  
##       cnt       
##  Min.   :  1.0  
##  1st Qu.: 40.0  
##  Median :142.0  
##  Mean   :189.5  
##  3rd Qu.:281.0  
##  Max.   :977.0

Based on the information from summary, there are no obvious issues.

Let’s also visually explore the data:

bikes %>% 
    select(-instant, -dteday) %>% 
    pivot_longer(everything()) %>% 
    ggplot() +
        aes(x = value) +
        geom_density() +
        facet_wrap(~name, scales = "free")

There are two things we can see here:

  • Some variables should be turned into factors, because while using numbers to represent their value, they are not continuous. In most cases this was already obvious from the description of the data, but we got a nice visual confirmation here. This is the case for variables like season or weekday with many clear defined peaks in the data distribution.
  • Windspeed has a suspicious peak at 0.

Let’s have a closer look at windspeed:

bikes %>% 
    select(windspeed) %>% 
    ggplot() +
        aes(x = windspeed) +
        geom_histogram(binwidth = 0.05)

There is an unexpected amount of 0 values in the data. It is difficult to say why this is the case, but likely those are essentially missings. Dropping the column might be reasonable.

Now, let’s take a closer look at our dependent variable: the amount of bikes rented:

bikes %>% 
    select(dteday, cnt, casual, registered) %>% 
    pivot_longer(-dteday) %>% 
    ggplot() + 
        aes(x = dteday, y = value) +
        geom_line() +
        facet_grid(name~.)

There is an obvious trend in the data for more rented bikes in 2012 compared to 2011. Under normal circumstances we should probably address that. Things we could do here are time series modelling and/or normalising our data by the total amount of registered users - if we were running a bike sharing business we should have that information available. However, this issue is beyond the scope of this post.

We will drop the casual and registered columns, as both are included in cnt. We can confirm this easily:

all(bikes$casual + bikes$registered == bikes$cnt)
## [1] TRUE

To finish the exploration we should also have a look at correlations between variables to identify pairs of highly correlated variables and also to get an idea of which predictors might be related to our dependent variable.

bikes %>% 
    select(-instant, -dteday) %>% 
    cor() %>% 
    as_tibble(rownames = "x") %>% 
    pivot_longer(-x) %>% 
    ggplot() +
        aes(x = x, y = name, fill = value) +
        geom_raster() +
        scale_fill_gradient2(low = "purple", mid = "white",
                             high = "orangered") +
        labs(x = NULL, y = NULL) +
        theme(axis.text.x = element_text(
            angle = 90, hjust = 1, vjust = 0.5))

From the correlation heat map we can see that temp and atemp are unsurprisingly highly correlated. Furthermore cnt appears to be to some extend correlated with time of the day and variables on weather conditions like temp and humidity.

So, after exploring the data, we will try dropping the windpseed column as it contains a suspicious amount of 0 values, as well as the casual and registered columns as we are only interested in the total amount if bikes rented. The temp and atemp variables are highly correlated, so we should remove one of them, but to explore tidymodels we will keep them for now and set up a preprocessing that handles those.

bikes2 <- bikes %>% 
    select(-casual, -registered, -windspeed)

We will also remove the year variable as we saw in the plot above that in 2012 much more bikes were rented than in 2011. While not exactly the best approach, as touched upon above, it should do well enough for demonstrating tidymodels.

bikes2 <- bikes2 %>% select(-yr)

As hr is discrete in our data and we will later dummy code our variables, we will try to reduce its levels here.

bikes2 <- bikes2 %>% 
  mutate(hr_cat = as_factor(hr)) %>% 
  mutate(hr_cat = fct_collapse(.$hr_cat,
    "night" = as.character(c(0:6)),
    "morning" = as.character(c(7:12)),
    "afternoon" = as.character(c(13:18)),
    "evening" = as.character(c(19:23))
  )) %>% 
  select(-hr)

Finally, we turn discrete variables into factors:

bikes2 <- bikes2 %>% 
  mutate_at(vars(season, mnth, holiday, weekday, 
                 workingday, weathersit), as_factor)

Preprocessing with tidymodels

The next thing you probably want to do is splitting the data into test and train dataset. We shuffle the rows of the dataset to not split at a time point when we divide the data. The default with sample_frac in dplyr is that it samples 100% of the data without replacement, giving a nice tool for quick shuffling of rows.

bikes2 <- sample_frac(bikes2)

We first prepare the split by calling initial_split from the rsample package from the tidymodels family. Here, we theoretically could add stratified splitting and of course set the amount of data to put into each split, but the default of using 75% of data is a sane choice.

bike_split <- initial_split(bikes2)
bike_split
## <13035/4344/17379>

The initial_split function returns a rset object defining the split, but not carrying it out yet.

Next, we extract our actual train and test datasets, which turn out to be tibbles.

bike_train <- training(bike_split)
bike_test <- testing(bike_split)
bike_train
## # A tibble: 13,035 x 13
##    instant dteday     season mnth  holiday weekday workingday weathersit  temp
##      <dbl> <date>     <fct>  <fct> <fct>   <fct>   <fct>      <fct>      <dbl>
##  1    3479 2011-05-29 2      5     0       0       0          1           0.7 
##  2    7754 2011-11-24 4      11    1       4       0          1           0.5 
##  3    8811 2012-01-07 1      1     0       6       0          1           0.44
##  4    9028 2012-01-17 1      1     0       2       1          2           0.26
##  5   13832 2012-08-04 3      8     0       6       0          1           0.86
##  6    7776 2011-11-25 4      11    0       5       1          1           0.52
##  7    5503 2011-08-22 3      8     0       1       1          1           0.68
##  8    5836 2011-09-05 3      9     1       1       0          1           0.74
##  9    8925 2012-01-12 1      1     0       4       1          1           0.46
## 10    4601 2011-07-15 3      7     0       5       1          1           0.7 
## # … with 13,025 more rows, and 4 more variables: atemp <dbl>, hum <dbl>,
## #   cnt <dbl>, hr_cat <fct>

Depending on the data and the methods you use, you often need to do some kind of preprocessing. This usually involves removal of highly correlated variables, dummy coding as well as scaling and centring of the data. Tidymodels allows you to define a pipeline for preprocessing you can then apply to the data, using tools from the recipes package.

To do so you first define what kind of variables you have:

base_recipe <- recipe(x = bike_train) %>%
  update_role(everything(), new_role = "predictor") %>% 
  update_role(instant, dteday, new_role = "idvar") %>% 
  update_role(cnt, new_role = "outcome")

Here we added roles to the variables in our data. This allows us to specify what we want to do with the data in our preprocessing pipeline as for everything we do in there we can define to which roles and variable types the steps should be applied.

We can choose the names of our roles pretty much the way we like them, however there are selection functions for predictor so you might want to use that role for your predictors.

What roles do should become clearer when we build and apply our pipeline.

pipeline <- base_recipe %>%
  step_nzv(all_predictors()) %>%
  step_knnimpute(all_predictors()) %>%
  step_center(all_numeric(), -has_role("outcome"),  -has_role("idvar")) %>%
  step_scale(all_numeric(), -has_role("outcome"),  -has_role("idvar")) %>%
  step_dummy(all_predictors(), -all_numeric()) %>%
  step_corr(all_numeric(), -has_role("outcome"),  -has_role("idvar"),
            threshold = 0.9)

Here we remove data with near zero variance, impute missing values using k-Nearest-Neighbours (there are no missings in our data, but I included this step for demonstrating it), dummy code categorical variables, centre and scale and finally remove variables if they highly correlate with another.

Using selector functions we specify to which role or type each step should be applied. With all_predictors we choose everything that has the role predictor, while with has_role we can select specific roles. If a - is added before a selector we exclude that type or role, allowing us to easily specify the set of variables we want a step to be applied to.

Our defined preprocessing pipeline can then be initialised, which does the estimation of parameters for the steps.

pipeline_prep <- prep(pipeline, training = bike_train,
                      strings_as_factors = FALSE)

We can also inspect the object to get some more information about the steps. For example we can see that the holiday column will be removed due to it having almost no variance and temp will be removed due to it being highly correlated with another variable (most likely atemp).

pipeline_prep
## Data Recipe
## 
## Inputs:
## 
##       role #variables
##      idvar          2
##    outcome          1
##  predictor         10
## 
## Training data contained 13035 data points and no missing data.
## 
## Operations:
## 
## Sparse, unbalanced variable filter removed holiday [trained]
## K-nearest neighbor imputation for mnth, weekday, workingday, weathersit, ... [trained]
## Centering for temp, atemp, hum [trained]
## Scaling for temp, atemp, hum [trained]
## Dummy variables from season, mnth, weekday, workingday, weathersit, hr_cat [trained]
## Correlation filter removed temp [trained]

And then we apply it to the data.

training <- bake(pipeline_prep, new_data = bike_train)
testing <- bake(pipeline_prep, new_data = bike_test)

Our training dataset now looks like this:

training
## # A tibble: 13,035 x 32
##    instant dteday       atemp     hum   cnt season_X2 season_X3 season_X4
##      <dbl> <date>       <dbl>   <dbl> <dbl>     <dbl>     <dbl>     <dbl>
##  1    3479 2011-05-29  1.12    0.846    219         1         0         0
##  2    7754 2011-11-24  0.0603 -1.48     114         0         0         1
##  3    8811 2012-01-07 -0.204  -1.27     109         0         0         0
##  4    9028 2012-01-17 -1.44    0.381     12         0         0         0
##  5   13832 2012-08-04  1.83   -1.17     547         0         1         0
##  6    7776 2011-11-25  0.149  -1.22     272         0         0         1
##  7    5503 2011-08-22  0.943   0.122     12         0         1         0
##  8    5836 2011-09-05  1.30    0.381    357         0         1         0
##  9    8925 2012-01-12 -0.116   0.0188   495         0         0         0
## 10    4601 2011-07-15  1.03   -0.602    265         0         1         0
## # … with 13,025 more rows, and 24 more variables: mnth_X2 <dbl>, mnth_X3 <dbl>,
## #   mnth_X4 <dbl>, mnth_X5 <dbl>, mnth_X6 <dbl>, mnth_X7 <dbl>, mnth_X8 <dbl>,
## #   mnth_X9 <dbl>, mnth_X10 <dbl>, mnth_X11 <dbl>, mnth_X12 <dbl>,
## #   weekday_X1 <dbl>, weekday_X2 <dbl>, weekday_X3 <dbl>, weekday_X4 <dbl>,
## #   weekday_X5 <dbl>, weekday_X6 <dbl>, workingday_X1 <dbl>,
## #   weathersit_X2 <dbl>, weathersit_X3 <dbl>, weathersit_X4 <dbl>,
## #   hr_cat_morning <dbl>, hr_cat_afternoon <dbl>, hr_cat_evening <dbl>

Under normal circumstances we should have dropped the idvar and dteday columns beforehand, but I wanted to demonstrate the use of roles in preprocessing with tidymodels and thus left them in.

training <- training %>% 
  select(-instant, -dteday)

Building and evaluating a model

Linear Regression

First, we go with a simple approach. When modelling the first thing you try should almost always be a linear model, so let’s build one:

lmod <- linear_reg() %>% 
  set_engine("lm")

When using tidymodels you create a model object and select an engine which it should use before fitting and predicting (to be more specific: those functions are provided by the parsnip package). In the case above we say that we want a linear regression using lm. Now, what’s the difference to just calling lm directly? It is the workflow: for a single model family seen in isolation it does not matter much, but when using several different model families you are quickly confronted with slight differences in using them. tidymodels however, enables you to use the exact same workflow for various kinds of models without having to worry about differences between different packages.

We can fit our model by passing it into the fit function and specifying the formula we want to use.

lmod_fit <- lmod %>% fit(training, formula = cnt ~ .)

And now we can make predictions. You can just pass the model into the predict function if you like, but here we want to combine the predictions with the data, so we will opt for a slightly more complex construction with mutate.

training2 <- training %>% 
  mutate(lm_pred = lmod_fit %>% 
           predict(training) %>% 
           deframe())
testing <- testing %>% 
  mutate(lm_pred = lmod_fit %>% 
           predict(testing) %>% 
           deframe())

Using the metrics function from the yardstick package we can get some metrics about how good our predictions are. First we check on the training data.

training2 %>% metrics(cnt, lm_pred)
## # A tibble: 3 x 3
##   .metric .estimator .estimate
##   <chr>   <chr>          <dbl>
## 1 rmse    standard     129.   
## 2 rsq     standard       0.500
## 3 mae     standard      95.3

And now on the test data.

testing %>% metrics(cnt, lm_pred)
## # A tibble: 3 x 3
##   .metric .estimator .estimate
##   <chr>   <chr>          <dbl>
## 1 rmse    standard     128.   
## 2 rsq     standard       0.495
## 3 mae     standard      94.0

Errors on train and test data are similar, suggesting that there is no overfitting.

To evaluate how well our model does work we can of course plot the predicted data against the actual data:

ggplot(testing) +
  aes(x = cnt, y = lm_pred) +
  geom_abline(intercept = 0, slope = 1) +
  geom_point(alpha = .5) 

The plot does not exactly suggest that our predictive performance is good. There are two obvious issues: 1. we predict values below 0 which does not make sense given the data; 2. we have really large prediction errors for large values of cnt.

You might also wonder how you access the underlying lm object. Well, this is easy:

ls(lmod_fit)
## [1] "elapsed" "fit"     "lvl"     "preproc" "spec"

The actual model is contained in the fit.

summary(lmod_fit$fit)
## 
## Call:
## stats::lm(formula = formula, data = data)
## 
## Residuals:
##     Min      1Q  Median      3Q     Max 
## -330.15  -80.56  -22.76   58.11  584.68 
## 
## Coefficients:
##                  Estimate Std. Error t value Pr(>|t|)    
## (Intercept)         7.317      6.491   1.127 0.259697    
## atemp              45.477      2.371  19.177  < 2e-16 ***
## hum               -18.518      1.461 -12.672  < 2e-16 ***
## season_X2          42.530      7.096   5.993 2.11e-09 ***
## season_X3          33.197      8.475   3.917 9.01e-05 ***
## season_X4          69.134      7.125   9.704  < 2e-16 ***
## mnth_X2             5.418      5.681   0.954 0.340273    
## mnth_X3             9.758      6.345   1.538 0.124107    
## mnth_X4            -2.812      9.459  -0.297 0.766221    
## mnth_X5            19.435     10.014   1.941 0.052313 .  
## mnth_X6             2.682     10.137   0.265 0.791298    
## mnth_X7           -14.798     11.499  -1.287 0.198144    
## mnth_X8            10.286     11.173   0.921 0.357266    
## mnth_X9            32.998     10.084   3.272 0.001069 ** 
## mnth_X10           10.740      9.398   1.143 0.253166    
## mnth_X11          -10.123      9.079  -1.115 0.264878    
## mnth_X12           -7.748      7.203  -1.076 0.282082    
## weekday_X1        -12.298      7.373  -1.668 0.095352 .  
## weekday_X2        -11.255      8.225  -1.368 0.171214    
## weekday_X3         -3.819      8.212  -0.465 0.641879    
## weekday_X4         -2.486      8.159  -0.305 0.760599    
## weekday_X5         -1.470      8.163  -0.180 0.857091    
## weekday_X6         16.059      4.225   3.801 0.000145 ***
## workingday_X1      20.735      7.106   2.918 0.003530 ** 
## weathersit_X2     -11.880      2.799  -4.244 2.21e-05 ***
## weathersit_X3     -70.077      4.596 -15.247  < 2e-16 ***
## weathersit_X4     111.541    128.945   0.865 0.387040    
## hr_cat_morning    189.851      3.205  59.235  < 2e-16 ***
## hr_cat_afternoon  248.584      3.713  66.956  < 2e-16 ***
## hr_cat_evening    133.024      3.398  39.153  < 2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 128.8 on 13005 degrees of freedom
## Multiple R-squared:  0.4999, Adjusted R-squared:  0.4988 
## F-statistic: 448.3 on 29 and 13005 DF,  p-value: < 2.2e-16

Random Forest

Next try something a little bit more flexible and try to mitigate the issues we have seen with our linear model. The workflow is pretty much the same. We first define our model.

rf <- rand_forest() %>%
  set_mode("regression") %>% 
  set_engine("ranger")

Then fit it.

rf_fit <- rf %>% fit(training, formula = cnt ~ .)

And now we can make predictions, again on both data sets.

training2 <- training %>% 
  mutate(rf_pred = rf_fit %>% 
           predict(training) %>% 
           deframe())

testing <- testing %>% 
  mutate(rf_pred = rf_fit %>% 
           predict(testing) %>% 
           deframe())

And of course we can also evaluate our model’s performance the same way as above.

training2 %>% metrics(cnt, rf_pred)
## # A tibble: 3 x 3
##   .metric .estimator .estimate
##   <chr>   <chr>          <dbl>
## 1 rmse    standard      97.8  
## 2 rsq     standard       0.736
## 3 mae     standard      69.0
testing %>% metrics(cnt, rf_pred)
## # A tibble: 3 x 3
##   .metric .estimator .estimate
##   <chr>   <chr>          <dbl>
## 1 rmse    standard     113.   
## 2 rsq     standard       0.614
## 3 mae     standard      79.3

It seems that just by choosing a different model we were able to reduce our prediction error, however at the same time we are now overfitting against the training data.

To wrap this section up, let’s also plot the predictions as generated with the random forest against the true values.

ggplot(testing) +
  aes(x = cnt, y = rf_pred) +
  geom_abline(intercept = 0, slope = 1) +
  geom_point(alpha = .5) 

This looks better than what we got with linear regression, but there is still room for improvement. The issue of large errors for large values of cnt is still present despite being less severe.

Cross Validation

Obviously we are not limited to a simple workflow of fitting models and predicting afterwards. Given that we are appearing to overfit with our random forest model, cross validation is an option we should explore, so we get a reasonable estimate of which metrics we can expect.

Cross validation in tidymodels is also a very simple process. The functions used here are part of the rsample and tune packages from the tidymodels family.

First we create our folds, ten in this case.

training_cv <- vfold_cv(training, v = 10)

This does create a list of splits similar to what we created with initial_split above. Carrying out the fitting and predicting on this list is equally simple.

rf_cv <- fit_resamples(
  formula = cnt ~ .,
  model = rf,
  resamples = training_cv
)

The formula is the usual definition of dependent variable and predictors, the model is our model specified above and the resamples are the folds we just created.

And we can have a look at the metrics generated here.

rf_cv %>% collect_metrics()
## # A tibble: 2 x 5
##   .metric .estimator    mean     n std_err
##   <chr>   <chr>        <dbl> <int>   <dbl>
## 1 rmse    standard   115.       10 0.700  
## 2 rsq     standard     0.607    10 0.00525

We can also get the range of RMSE values doing a bit of wrangling with the results.

rf_cv %>% 
  unnest(.metrics) %>% 
  filter(.metric == "rmse") %>% 
  select(.estimate) %>% 
  deframe() %>% 
  range()
## [1] 111.9653 119.9293

The values we got here are a better estimate on how the model would perform on previously unseen data and this is comparable to what we have seen on our test data.

Model tuning

With many kinds of models we have the option to tune hyperparameters, this is where the tools provided by the tune package come in. Hyperparameter tuning is often not an exact science but boils down to carrying out a grid search, i.e. trying a pre-defined range of values for a set of hyperparameters and see how the model performs.

Most likely you want to tune more than one parameter, but for demonstration purposes we will just tune the mtry parameter for rand_forest, controlling the amount of predictors to sample at each tree split.

rf_grid <- rand_forest(mtry = tune()) %>% 
  set_mode("regression") %>% 
  set_engine("ranger")

We can then conduct the grid search. Be aware that doing this might take quite some time, depending on the model complexity and the amount of data. You can set the control parameter verbose to TRUE to get updates about the status.

rf_grid_fit <- tune_grid(
  formula = cnt ~ .,
  model = rf_grid, 
  resamples = training_cv,
  control = control_grid(verbose = FALSE) 
)

You can then extract the best parameters from the fitted grid. Make sure to specify whether you want to maximize the metric of interest, with RMSE you obviously want the lowest value.

rf_grid_fit %>% show_best(metric = "rmse", maximize = FALSE)
## # A tibble: 5 x 6
##    mtry .metric .estimator  mean     n std_err
##   <int> <chr>   <chr>      <dbl> <int>   <dbl>
## 1     9 rmse    standard    112.    10   0.858
## 2    12 rmse    standard    112.    10   0.873
## 3    13 rmse    standard    112.    10   0.874
## 4    15 rmse    standard    112.    10   0.841
## 5    18 rmse    standard    113.    10   0.805

Apparently an mtry value of 9 provides the lowest error, but comparing it to the other values here we are clearly in the range of diminishing returns for the given circumstances.

Conclusions

The tidymodels family of packages sure looks promising. This post has only shown the tip of the iceberg regarding what you can do, however this is hopefully good enough to get you up to speed.

As of now you probably want to be careful putting tidymodels to productive use as those packages have not matured yet and things are likely changing quite significantly in the future, so stuff might break when updating packages.

Additionally one thing that has not made it into tidymodels yet is mixed modelling, which is something that might be essential for you. There is the option to add your own model interfaces, but I have not given this a try yet and thus am unsure whether this is suitable for mixed models.

To sum things up: tidymodels is likely here to stay and you might want to familiarise yourself with it, but do not fully rely on it yet.