tidypredict

The main goal of tidypredict is to enable running predictions inside databases. It reads the model, extracts the components needed to calculate the prediction, and then creates an R formula that can be translated into SQL. In other words, it is able to parse a model such as this one:

model <- lm(mpg ~ wt + cyl, data = mtcars)

tidypredict can return a SQL statement that is ready to run inside the database. Because it uses dplyr’s database interface, it works with several databases back-ends, such as MS SQL:

tidypredict_sql(model, dbplyr::simulate_mssql())

## <SQL> 39.6862614802529 + (`wt` * -3.19097213898374) + (`cyl` * -1.5077949682598)

Installation

Install tidypredict from CRAN using:

# install.packages("tidypredict")

Or install the development version using devtools as follows:

# install.packages("remotes")
# remotes::install_github("tidymodels/tidypredict")

Functions

tidypredict has only a few functions, and it is not expected that number to grow much. The main focus at this time is to add more models to support.

How it works

Instead of translating directly to a SQL statement, tidypredict creates an R formula. That formula can then be used inside dplyr. The overall workflow would be as illustrated in the image above, and described here:

1. Fit the model using a base R model, or one from the packages listed in Supported Models
2. tidypredict reads model, and creates a list object with the necessary components to run predictions
3. tidypredict builds an R formula based on the list object
4. dplyr evaluates the formula created by tidypredict
5. dplyr translates the formula into a SQL statement, or any other interfaces.
6. The database executes the SQL statement(s) created by dplyr

Parsed model spec

tidypredict writes and reads a spec based on a model. Instead of simply writing the R formula directly, splitting the spec from the formula adds the following capabilities:

1. No more saving models as .rds - Specifically for cases when the model needs to be used for predictions in a Shiny app.
2. Beyond R models - Technically, anything that can write a proper spec, can be read into tidypredict. It also means, that the parsed model spec can become a good alternative to using PMML.

Supported models

The following models are supported by tidypredict:

• Linear Regression - lm()
• Generalized Linear model - glm()
• Random Forest models - randomForest::randomForest()
• Random Forest models, via ranger - ranger::ranger()
• MARS models - earth::earth()
• XGBoost models - xgboost::xgb.Booster.complete()
• Cubist models - Cubist::cubist()
• Tree models, via partykit - partykit::ctree()

parsnip

tidypredict supports models fitted via the parsnip interface. The ones confirmed currently work in tidypredict are:

• lm() - parsnip: linear_reg() with “lm” as the engine.
• randomForest::randomForest() - parsnip: rand_forest() with “randomForest” as the engine.
• ranger::ranger() - parsnip: rand_forest() with “ranger” as the engine.
• earth::earth() - parsnip: mars() with “earth” as the engine.

broom

The tidy() function from broom works with linear models parsed via tidypredict

pm <- parse_model(lm(wt ~ ., mtcars))
tidy(pm)

## # A tibble: 11 x 2
##    term        estimate
##    <chr>          <dbl>
##  1 (Intercept) -0.231  
##  2 mpg         -0.0417 
##  3 cyl         -0.0573 
##  4 disp         0.00669
##  5 hp          -0.00323
##  6 drat        -0.0901 
##  7 qsec         0.200  
##  8 vs          -0.0664 
##  9 am           0.0184 
## 10 gear        -0.0935 
## 11 carb         0.249