In brunocarlin/tidy.outliers: What the Package Does (One Line, Title Case)

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "man/figures/README-",
  out.width = "100%"
)

tidy.outliers

The goal of tidy.outliers is to allow for easy usage of many outliers removal methods, currently implemented are:

Simple methods:

• Univariate based function
• Mahalanobis distance

Model Methods:

• lookout
• h2o.extendedIsolationForest
• outForest

What are outlier scores?

The package works on the principal that all basic step_outlier_* functions return an outlier "score" that can be used for filtering outliers where 0 is a very low outlier score and 1 is a very high outlier score, so you could filter, for example all rows where the outlier score is greater than .9.

Installation

You can not yet install the released version of tidy.outliers from CRAN with:

#install.packages("tidy.outliers")

And the development version from GitHub with:

# install.packages("devtools")
devtools::install_github("brunocarlin/tidy.outliers")

Usage

Load Libraries

library(recipes)
library(tidy.outliers)

Create a recipe for calculation the outlier scores

I keep the mpg as an example outcome since you should remove outlier from your outcome, you also shouldn't remove outlier from testing data so the default is to skip the steps of the package when predicting.

rec_obj <-
  recipe(mpg ~ ., data = mtcars) |>
  step_outliers_maha(all_numeric(), -all_outcomes()) |>
  step_outliers_lookout(all_numeric(),-contains(r"(.outliers)"),-all_outcomes()) |> 
  prep(mtcars)

Return scores

bake(rec_obj,new_data = NULL) |> 
  select(contains(r"(.outliers)")) |> 
  arrange(.outliers_lookout |> desc())

Example filtering based on scores

Create recipe filtering outliers

rec_obj2 <-
  recipe(mpg ~ ., data = mtcars) |>
  step_outliers_maha(all_numeric(), -all_outcomes()) |>
  step_outliers_lookout(all_numeric(),-contains(r"(.outliers)"),-all_outcomes()) |> 
  step_outliers_remove(contains(r"(.outliers)")) |> 
  prep(mtcars)

Returns the filtered rows

We filtered one row from the dataset and and automatically removed the extra outlier columns.

bake(rec_obj2,new_data = NULL) |> glimpse()

mtcars |> glimpse()

Investigate why some rows were filtered

And we can get which were the outliers and their score

tidy(rec_obj2,number = 3) |> 
  arrange(aggregation_results |> desc())

Integration with tidymodels

The package was made to play nice with tune and friends from tidymodels check out the article on our github pkgdown page!

Next steps

Although it is possible to manually change the function of model parameters using the options argument it would be nice to add the option to tune those internal parameters as well.

So instead of this.

rec_obj2 <-
  recipe(mpg ~ ., data = mtcars) |> 
  step_outliers_outForest(
    all_numeric(),
    -all_outcomes(),
    options = list(
    impute_multivariate_control = list(
      num.trees = 200
    )
  ))

You would write something like this

rec_obj2 <-
  recipe(mpg ~ ., data = mtcars) |> 
  step_outliers_outForest(
    all_numeric(),
    -all_outcomes(),
    options = list(
    impute_multivariate_control = list(
      num.trees = tune::tune('tree')
    )
  ))

The main problem is that this would require manually going model by model and incorporating those arguments as tunable components.

brunocarlin/tidy.outliers documentation built on Feb. 22, 2023, 8:24 p.m.