groupdata2

R package: Subsetting Methods for Balanced Cross-Validation, Time Series Windowing, and General Grouping and Splitting of Data.

By Ludvig R. Olsen Cognitive Science, Aarhus University Started in Oct. 2016

Contact at: r-pkgs@ludvigolsen.dk

Main functions:

Other tools:

Installation

Development version:

install.packages("devtools")

devtools::install_github("LudvigOlsen/groupdata2")

To do

Functions

group_factor()

Returns a factor with group numbers, e.g. (1,1,1,2,2,2,3,3,3).

This can be used to subset, aggregate, group_by, etc.

Create equally sized groups by setting force_equal = TRUE

Randomize grouping factor by setting randomize = TRUE

group()

Returns the given data as a dataframe with added grouping factor made with group_factor(). The dataframe is grouped by the grouping factor for easy use with dplyr pipelines.

splt()

Creates the specified groups with group_factor() and splits the given data by the grouping factor with base::split. Returns the splits in a list.

fold()

Creates (optionally) balanced folds for use in cross-validation. Balance folds on one categorical variable and/or make sure that all datapoints sharing an ID is in the same fold.

Methods

There are currently 6 methods available. They can be divided into 3 categories.

Examples of group sizes are based on a vector with 57 elements.

Specify group size

Method: greedy

Divides up the data greedily given a specified group size.

E.g. group sizes: 10, 10, 10, 10, 10, 7

Specify number of groups

Method: n_dist (Default)

Divides the data into a specified number of groups and distributes excess data points across groups.

E.g. group sizes: 11, 11, 12, 11, 12

Method: n_fill

Divides the data into a specified number of groups and fills up groups with excess data points from the beginning.

E.g. group sizes: 12, 12, 11, 11, 11

Method: n_last

Divides the data into a specified number of groups. The algorithm finds the most equal group sizes possible, using all data points. Only the last group is able to differ in size.

E.g. group sizes: 11, 11, 11, 11, 13

Method: n_rand

Divides the data into a specified number of groups. Excess data points are placed randomly in groups (only 1 per group).

E.g. group sizes: 12, 11, 11, 11, 12

Specify step size

Method: staircase

Uses step_size to divide up the data. Group size increases with 1 step for every group, until there is no more data.

E.g. group sizes: 5, 10, 15, 20, 7

Examples

# Attach packages
library(groupdata2)
library(dplyr)
library(knitr)
# Create dataframe
df <- data.frame("x"=c(1:12),
  "species" = rep(c('cat','pig', 'human'), 4),
  "age" = sample(c(1:100), 12))

group()

# Using group()
group(df, 5, method = 'n_dist') %>%
  kable()
x species age .groups
1 cat 44 1
2 pig 55 1
3 human 13 2
4 cat 1 2
5 pig 28 3
6 human 20 3
7 cat 56 3
8 pig 99 4
9 human 14 4
10 cat 47 5
11 pig 86 5
12 human 82 5 

# Using group() with dplyr pipeline to get mean age
df %>%
  group(5, method = 'n_dist') %>%
  dplyr::summarise(mean_age = mean(age)) %>%
  kable()
.groups mean_age
1 49.50000
2 7.00000
3 34.66667
4 56.50000
5 71.66667

fold()

# Create dataframe
df <- data.frame(
  "participant" = factor(rep(c('1','2', '3', '4', '5', '6'), 3)),
  "age" = rep(c(20,23,27,21,32,31), 3),
  "diagnosis" = rep(c('a', 'b', 'a', 'b', 'b', 'a'), 3),
  "score" = c(10,24,15,35,24,14,24,40,30,50,54,25,45,67,40,78,62,30))
df <- df[order(df$participant),]
df$session <- rep(c('1','2', '3'), 6)
# Using fold() 

# First set seed to ensure reproducibility
set.seed(1)

# Use fold() with cat_col and id_col
df_folded <- fold(df, 3, cat_col = 'diagnosis',
                  id_col = 'participant', method = 'n_dist')

# Show df_folded ordered by folds
df_folded[order(df_folded$.folds),] %>%
  kable()
participant age diagnosis score session .folds
1 20 a 10 1 1
1 20 a 24 2 1
1 20 a 45 3 1
4 21 b 35 1 1
4 21 b 50 2 1
4 21 b 78 3 1
6 31 a 14 1 2
6 31 a 25 2 2
6 31 a 30 3 2
5 32 b 24 1 2
5 32 b 54 2 2
5 32 b 62 3 2
3 27 a 15 1 3
3 27 a 30 2 3
3 27 a 40 3 3
2 23 b 24 1 3
2 23 b 40 2 3
2 23 b 67 3 3 

# Show distribution of diagnoses and participants
df_folded %>% 
  group_by(.folds) %>% 
  count(diagnosis, participant) %>% 
  kable()
.folds diagnosis participant n
1 a 1 3
1 b 4 3
2 a 6 3
2 b 5 3
3 a 3 3
3 b 2 3

Notice that the we now have the opportunity to include the session variable and/or use participant as a random effect in our model when doing cross-validation, as any participant will only appear in one fold.

We also have a balance in the representation of each diagnosis, which could give us better, more consistent results.