Automatic Machine Learning with 'tidymodels'.
tidyAML 
Introduction
Welcome to {tidyAML} which is a new R package that makes it easy to use the tidymodels ecosystem to perform automated machine learning (AutoML). This package provides a simple and intuitive interface that allows users to quickly generate machine learning models without worrying about the underlying details. It also includes a safety mechanism that ensures that the package will fail gracefully if any required extension packages are not installed on the user’s machine. With {tidyAML}, users can easily build high-quality machine learning models in just a few lines of code. Whether you are a beginner or an experienced machine learning practitioner, {tidyAML} has something to offer.
Some ideas are that we should be able to generate regression models on the fly without having to actually go through the process of building the specification, especially if it is a non-tuning model, meaning we are not planing on tuning hyper-parameters like penalty and cost.
The idea is not to re-write the excellent work the tidymodels team has done (because it’s not possible) but rather to try and make an enhanced easy to use set of functions that do what they say and can generate many models and predictions at once.
This is similar to the great h2o package, but, {tidyAML} does not require java to be setup properly like h2o because {tidyAML} is built on tidymodels.
Thanks
Thank you Garrick Aden-Buie for the easy name change suggestion.
Installation
You can install {tidyAML} like so:
#install.packages("tidyAML")
Or the development version from GitHub
# install.packages("devtools")
#devtools::install_github("spsanderson/tidyAML")
Examples
Part of the reason to use {tidyAML} is so that you can generate many models of your data set. One way of modeling a data set is using regression for some numeric output. There is a convienent function in tidyAML that will generate a set of non-tuning models for fast regression. Let’s take a look below.
First let’s load the library
library(tidyAML)
#> Loading required package: parsnip
#>
#> == Welcome to tidyAML ===========================================================================
#> If you find this package useful, please leave a star:
#> https://github.com/spsanderson/tidyAML'
#>
#> If you encounter a bug or want to request an enhancement please file an issue at:
#> https://github.com/spsanderson/tidyAML/issues
#>
#> It is suggested that you run tidymodels::tidymodel_prefer() to set the defaults for your session.
#>
#> Thank you for using tidyAML!
Now lets see the function in action.
fast_regression_parsnip_spec_tbl(.parsnip_fns = "linear_reg")
#> # A tibble: 11 × 5
#> .model_id .parsnip_engine .parsnip_mode .parsnip_fns model_spec
#> <int> <chr> <chr> <chr> <list>
#> 1 1 lm regression linear_reg <spec[+]>
#> 2 2 brulee regression linear_reg <spec[+]>
#> 3 3 gee regression linear_reg <spec[+]>
#> 4 4 glm regression linear_reg <spec[+]>
#> 5 5 glmer regression linear_reg <spec[+]>
#> 6 6 glmnet regression linear_reg <spec[+]>
#> 7 7 gls regression linear_reg <spec[+]>
#> 8 8 lme regression linear_reg <spec[+]>
#> 9 9 lmer regression linear_reg <spec[+]>
#> 10 10 stan regression linear_reg <spec[+]>
#> 11 11 stan_glmer regression linear_reg <spec[+]>
fast_regression_parsnip_spec_tbl(.parsnip_eng = c("lm","glm"))
#> # A tibble: 3 × 5
#> .model_id .parsnip_engine .parsnip_mode .parsnip_fns model_spec
#> <int> <chr> <chr> <chr> <list>
#> 1 1 lm regression linear_reg <spec[+]>
#> 2 2 glm regression linear_reg <spec[+]>
#> 3 3 glm regression poisson_reg <spec[+]>
fast_regression_parsnip_spec_tbl(.parsnip_eng = c("lm","glm","gee"),
.parsnip_fns = "linear_reg")
#> # A tibble: 3 × 5
#> .model_id .parsnip_engine .parsnip_mode .parsnip_fns model_spec
#> <int> <chr> <chr> <chr> <list>
#> 1 1 lm regression linear_reg <spec[+]>
#> 2 2 gee regression linear_reg <spec[+]>
#> 3 3 glm regression linear_reg <spec[+]>
As shown we can easily select the models we want either by choosing the supported parsnip function like linear_reg() or by choose the desired engine, you can also use them both in conjunction with each other!
This function also does add a class to the output. Let’s see it.
class(fast_regression_parsnip_spec_tbl())
#> [1] "tidyaml_mod_spec_tbl" "fst_reg_spec_tbl" "tidyaml_base_tbl"
#> [4] "tbl_df" "tbl" "data.frame"
We see that there are two added classes, first fst_reg_spec_tbl because this creates a set of non-tuning regression models and then tidyaml_mod_spec_tbl because this is a model specification tibble built with {tidyAML}
Now, what if you want to create a non-tuning model spec without using the fast_regression_parsnip_spec_tbl() function. Well, you can. The function is called create_model_spec().
create_model_spec(
.parsnip_eng = list("lm","glm","glmnet","cubist"),
.parsnip_fns = list(
"linear_reg",
"linear_reg",
"linear_reg",
"cubist_rules"
)
)
#> # A tibble: 4 × 4
#> .parsnip_engine .parsnip_mode .parsnip_fns .model_spec
#> <chr> <chr> <chr> <list>
#> 1 lm regression linear_reg <spec[+]>
#> 2 glm regression linear_reg <spec[+]>
#> 3 glmnet regression linear_reg <spec[+]>
#> 4 cubist regression cubist_rules <spec[+]>
create_model_spec(
.parsnip_eng = list("lm","glm","glmnet","cubist"),
.parsnip_fns = list(
"linear_reg",
"linear_reg",
"linear_reg",
"cubist_rules"
),
.return_tibble = FALSE
)
#> $.parsnip_engine
#> $.parsnip_engine[[1]]
#> [1] "lm"
#>
#> $.parsnip_engine[[2]]
#> [1] "glm"
#>
#> $.parsnip_engine[[3]]
#> [1] "glmnet"
#>
#> $.parsnip_engine[[4]]
#> [1] "cubist"
#>
#>
#> $.parsnip_mode
#> $.parsnip_mode[[1]]
#> [1] "regression"
#>
#>
#> $.parsnip_fns
#> $.parsnip_fns[[1]]
#> [1] "linear_reg"
#>
#> $.parsnip_fns[[2]]
#> [1] "linear_reg"
#>
#> $.parsnip_fns[[3]]
#> [1] "linear_reg"
#>
#> $.parsnip_fns[[4]]
#> [1] "cubist_rules"
#>
#>
#> $.model_spec
#> $.model_spec[[1]]
#> Linear Regression Model Specification (regression)
#>
#> Computational engine: lm
#>
#>
#> $.model_spec[[2]]
#> Linear Regression Model Specification (regression)
#>
#> Computational engine: glm
#>
#>
#> $.model_spec[[3]]
#> Linear Regression Model Specification (regression)
#>
#> Computational engine: glmnet
#>
#>
#> $.model_spec[[4]]
#> Cubist Model Specification (regression)
#>
#> Computational engine: cubist
Now the reason we are here. Let’s take a look at the first function for modeling with {tidyAML}, fast_regression().
library(recipes)
library(dplyr)
rec_obj <- recipe(mpg ~ ., data = mtcars)
frt_tbl <- fast_regression(
.data = mtcars,
.rec_obj = rec_obj,
.parsnip_eng = c("lm","glm","gee"),
.parsnip_fns = "linear_reg",
.drop_na = FALSE
)
glimpse(frt_tbl)
#> Rows: 3
#> Columns: 8
#> $ .model_id <int> 1, 2, 3
#> $ .parsnip_engine <chr> "lm", "gee", "glm"
#> $ .parsnip_mode <chr> "regression", "regression", "regression"
#> $ .parsnip_fns <chr> "linear_reg", "linear_reg", "linear_reg"
#> $ model_spec <list> [~NULL, ~NULL, NULL, regression, TRUE, NULL, lm, TRUE]…
#> $ wflw <list> [cyl, disp, hp, drat, wt, qsec, vs, am, gear, carb, mp…
#> $ fitted_wflw <list> [cyl, disp, hp, drat, wt, qsec, vs, am, gear, carb, mp…
#> $ pred_wflw <list> [<tbl_df[64 x 3]>], <NULL>, [<tbl_df[64 x 3]>]
As we see above, one of the models has gracefully failed, thanks in part to the function purrr::safely(), which was used to make what I call safe_make functions.
Let’s look at the fitted workflow predictions.
frt_tbl$pred_wflw
#> [[1]]
#> # A tibble: 64 × 3
#> .data_category .data_type .value
#> <chr> <chr> <dbl>
#> 1 actual actual 15.2
#> 2 actual actual 19.2
#> 3 actual actual 22.8
#> 4 actual actual 33.9
#> 5 actual actual 26
#> 6 actual actual 19.2
#> 7 actual actual 15
#> 8 actual actual 27.3
#> 9 actual actual 24.4
#> 10 actual actual 17.3
#> # ℹ 54 more rows
#>
#> [[2]]
#> NULL
#>
#> [[3]]
#> # A tibble: 64 × 3
#> .data_category .data_type .value
#> <chr> <chr> <dbl>
#> 1 actual actual 15.2
#> 2 actual actual 19.2
#> 3 actual actual 22.8
#> 4 actual actual 33.9
#> 5 actual actual 26
#> 6 actual actual 19.2
#> 7 actual actual 15
#> 8 actual actual 27.3
#> 9 actual actual 24.4
#> 10 actual actual 17.3
#> # ℹ 54 more rows
Now let’s load the multilevelmod library so that we can run the gee linear regression.
library(multilevelmod)
rec_obj <- recipe(mpg ~ ., data = mtcars)
frt_tbl <- fast_regression(
.data = mtcars,
.rec_obj = rec_obj,
.parsnip_eng = c("lm","glm","gee"),
.parsnip_fns = "linear_reg"
)
extract_wflw_pred(frt_tbl, 1:3)
#> # A tibble: 192 × 4
#> .model_type .data_category .data_type .value
#> <chr> <chr> <chr> <dbl>
#> 1 lm - linear_reg actual actual 32.4
#> 2 lm - linear_reg actual actual 14.3
#> 3 lm - linear_reg actual actual 15.8
#> 4 lm - linear_reg actual actual 30.4
#> 5 lm - linear_reg actual actual 24.4
#> 6 lm - linear_reg actual actual 15
#> 7 lm - linear_reg actual actual 33.9
#> 8 lm - linear_reg actual actual 22.8
#> 9 lm - linear_reg actual actual 19.2
#> 10 lm - linear_reg actual actual 21.4
#> # ℹ 182 more rows
Getting Regression Residuals
Getting residuals is easy with {tidyAML}. Let’s take a look.
extract_regression_residuals(frt_tbl)
#> [[1]]
#> # A tibble: 32 × 4
#> .model_type .actual .predicted .resid
#> <chr> <dbl> <dbl> <dbl>
#> 1 lm - linear_reg 32.4 27.5 4.94
#> 2 lm - linear_reg 14.3 14.2 0.121
#> 3 lm - linear_reg 15.8 18.5 -2.71
#> 4 lm - linear_reg 30.4 30.6 -0.178
#> 5 lm - linear_reg 24.4 22.6 1.82
#> 6 lm - linear_reg 15 13.3 1.69
#> 7 lm - linear_reg 33.9 29.3 4.64
#> 8 lm - linear_reg 22.8 25.3 -2.53
#> 9 lm - linear_reg 19.2 17.6 1.62
#> 10 lm - linear_reg 21.4 21.2 0.162
#> # ℹ 22 more rows
#>
#> [[2]]
#> # A tibble: 32 × 4
#> .model_type .actual .predicted .resid
#> <chr> <dbl> <dbl> <dbl>
#> 1 gee - linear_reg 32.4 27.5 4.95
#> 2 gee - linear_reg 14.3 14.2 0.0928
#> 3 gee - linear_reg 15.8 18.5 -2.67
#> 4 gee - linear_reg 30.4 30.5 -0.147
#> 5 gee - linear_reg 24.4 22.6 1.83
#> 6 gee - linear_reg 15 13.3 1.68
#> 7 gee - linear_reg 33.9 29.3 4.65
#> 8 gee - linear_reg 22.8 25.3 -2.53
#> 9 gee - linear_reg 19.2 17.6 1.60
#> 10 gee - linear_reg 21.4 21.2 0.165
#> # ℹ 22 more rows
#>
#> [[3]]
#> # A tibble: 32 × 4
#> .model_type .actual .predicted .resid
#> <chr> <dbl> <dbl> <dbl>
#> 1 glm - linear_reg 32.4 27.5 4.94
#> 2 glm - linear_reg 14.3 14.2 0.121
#> 3 glm - linear_reg 15.8 18.5 -2.71
#> 4 glm - linear_reg 30.4 30.6 -0.178
#> 5 glm - linear_reg 24.4 22.6 1.82
#> 6 glm - linear_reg 15 13.3 1.69
#> 7 glm - linear_reg 33.9 29.3 4.64
#> 8 glm - linear_reg 22.8 25.3 -2.53
#> 9 glm - linear_reg 19.2 17.6 1.62
#> 10 glm - linear_reg 21.4 21.2 0.162
#> # ℹ 22 more rows
You can also pivot them into a long format making plotting easy with ggplot2.
extract_regression_residuals(frt_tbl, .pivot_long = TRUE)
#> [[1]]
#> # A tibble: 96 × 3
#> .model_type name value
#> <chr> <chr> <dbl>
#> 1 lm - linear_reg .actual 32.4
#> 2 lm - linear_reg .predicted 27.5
#> 3 lm - linear_reg .resid 4.94
#> 4 lm - linear_reg .actual 14.3
#> 5 lm - linear_reg .predicted 14.2
#> 6 lm - linear_reg .resid 0.121
#> 7 lm - linear_reg .actual 15.8
#> 8 lm - linear_reg .predicted 18.5
#> 9 lm - linear_reg .resid -2.71
#> 10 lm - linear_reg .actual 30.4
#> # ℹ 86 more rows
#>
#> [[2]]
#> # A tibble: 96 × 3
#> .model_type name value
#> <chr> <chr> <dbl>
#> 1 gee - linear_reg .actual 32.4
#> 2 gee - linear_reg .predicted 27.5
#> 3 gee - linear_reg .resid 4.95
#> 4 gee - linear_reg .actual 14.3
#> 5 gee - linear_reg .predicted 14.2
#> 6 gee - linear_reg .resid 0.0928
#> 7 gee - linear_reg .actual 15.8
#> 8 gee - linear_reg .predicted 18.5
#> 9 gee - linear_reg .resid -2.67
#> 10 gee - linear_reg .actual 30.4
#> # ℹ 86 more rows
#>
#> [[3]]
#> # A tibble: 96 × 3
#> .model_type name value
#> <chr> <chr> <dbl>
#> 1 glm - linear_reg .actual 32.4
#> 2 glm - linear_reg .predicted 27.5
#> 3 glm - linear_reg .resid 4.94
#> 4 glm - linear_reg .actual 14.3
#> 5 glm - linear_reg .predicted 14.2
#> 6 glm - linear_reg .resid 0.121
#> 7 glm - linear_reg .actual 15.8
#> 8 glm - linear_reg .predicted 18.5
#> 9 glm - linear_reg .resid -2.71
#> 10 glm - linear_reg .actual 30.4
#> # ℹ 86 more rows