featuretoolsR

An R interface to the Python module Featuretools.

General

featuretoolsR provides functionality from the Python module featuretools, which aims to automate feature engineering. This package is very much a work in progress as Featuretools offers a lot of functionality. Any PRs are much appreciated.

Installing

Package

CRAN

The latest stable release is found on CRAN.

Github

You can get the latest version of featuretoolsR by installing it straight from Github: devtools::install_github("magnusfurugard/featuretoolsR").

Featuretools

You'll need to have a working Python environment as well as featuretools installed. The recommended way is to use the built-in function install_featuretools() which automatically sets up a virtual environment for the package and installs featuretools.

Usage

All functions in featuretoolsR comes with documentation, but it's advised to briefly browse through the Featuretools Python documentation. It'll cover things like entities, relationships and dfs.

Creating an entityset

An entityset is the set which contain all your entities. To create a set and add an entity straight away, you can use as_entityset.

# Libs
library(featuretoolsR)
library(magrittr)

# Create some mock data
set_1 <- data.frame(key = 1:100, value = sample(letters, 100, T), a = rep(Sys.Date(), 100))
set_2 <- data.frame(key = 1:100, value = sample(LETTERS, 100, T), b = rep(Sys.time(), 100))

# Create entityset
es <- as_entityset(
  set_1, 
  index = "key", 
  entity_id = "set_1", 
  id = "demo", 
  time_index = "a"
)

Adding entities

To add entities (i.e if you have relational data across multiple data.frames), this can be achieved with add_entity. This function is pipe friendly. For this demo-case, we'll use set_2.

es <- es %>%
  add_entity(
    df = set_2, 
    entity_id = "set_2", 
    index = "key", 
    time_index = "b"
  )

Defining relationships

With relational data, it's useful to define a relationship between two or more entities. This can be done with add_relationship.

es <- es %>%
  add_relationship(
    parent_set = "set_1", 
    child_set = "set_2", 
    parent_idx = "key", 
    child_idx = "key"
  )

Deep feature synthesis

The bread and butter of Featuretools is the dfs-function (official docs here). It will attempt to create features based on *_primitives you provide (more on primitives below).

ft_matrix <- es %>%
  dfs(
    target_entity = "set_1", 
    trans_primitives = c("and", "cum_sum")
  )

Tidying up

To use the new data.frame/features created by dfs, a function unique for featuretoolsR, tidy_feature_matrix can be used. A few "nice-to-have" arguments can be passed to clean the new data, like removing near zero variance variables, as well as replacing NaN with NA.

tidy <- tidy_feature_matrix(ft_matrix, remove_nzv = T, nan_is_na = T, clean_names = T)

Primitives

Featuretools supports a lot of primitives. These are accessible with the function list_primitives() which returns a data.frame containing type (aggregation (agg_primitives) or transform (trans_primitives)), name (in the example above, "and" and "divide") as well as a brief description of the primitive itself.

Credits

reticulate - an R interface to Python.

Featuretools