knitr::opts_chunk$set( collapse = TRUE, comment = "#>", fig.align='center', dpi = 92, fig.retina = 2 ) options(tibble.print_min = 4L, tibble.print_max = 4L)
You can either install the CRAN version or the GitHub development version.
# Uncomment: # install.packages("groupdata2")
# Uncomment: # install.packages("devtools") # devtools::install_github("LudvigOlsen/groupdata2")
# Attaching groupdata2 library(groupdata2) # Attaching other packages used in this vignette library(dplyr) library(tidyr) require(ggplot2, quietly = TRUE) # Attach if installed library(knitr) # We will also be using plyr a few times, but we don't attach this # because of possible conflicts with dplyr. Instead we use its functions # like so: plyr::count()
groupdata2
is a set of functions and methods for easy grouping, windowing, folding, partitioning, splitting and balancing of data.
There are 7 main functions:
Returns a factor with group numbers, e.g. c(1, 1, 1, 2, 2, 2, 3, 3, 3)
.
This can be used to subset, aggregate, group_by, etc.
Returns the given data as a data frame
with an added grouping factor made with group_factor()
. The data frame
is grouped by the grouping factor for easy use with %>%
pipelines.
Splits the given data into the specified groups made with group_factor()
and returns them in a list
.
Creates (optionally) balanced folds for use in cross-validation. Balance folds on one categorical variable and/or one numerical variable. Ensure that all datapoints sharing an ID is in the same fold. Can create multiple unique fold columns at once, e.g. for repeated cross-validation.
Creates (optionally) balanced partitions (e.g. training/test sets). Balance partitions on one categorical variable and/or one numerical variable. Make sure that all datapoints sharing an ID is in the same partition.
Collapses existing groups into a smaller set of groups with (optional) categorical, numerical, ID, and/or size balancing.
Uses up- or downsampling to fix the size of all groups to the min, max, mean, or median group size or to a specific number of rows. Balancing can also happen on the ID level, e.g. to ensure the same number of IDs in each category.
When working with time series we would often refer to the kind of groups made by group_factor()
, group()
and splt()
as windows.
In this vignette, these will be referred to as groups.
fold()
creates balanced groups for cross-validation by using group()
. These are referred to as folds.
partition()
creates balanced groups of (potentially) different sizes, by using group()
. These are referred to as partitions.
Cross-validation with groupdata2
In this vignette, we go through the basics of cross-validation, such as creating balanced train/test sets with partition()
and balanced folds with fold()
. We also write up a simple cross-validation function and compare multiple linear regression models.
Time series with groupdata2
In this vignette, we divide up a time series into groups (windows) and subgroups using group()
with the greedy
and staircase
methods. We do some basic descriptive stats of each group and use them to reduce the data size.
Automatic groups with groupdata2
In this vignette, we will use the l_starts
method with group()
to allow transferring of information from one dataset to another. We will use the automatic grouping function that finds group starts all by itself.
In the examples, we will be using knitr::kable()
to visualize some of the data such as data frames. You do not need to use kable()
when using the functions in groupdata2
.
There are currently 10 methods for grouping the data.
It is possible to create groups based on number of groups (default), group size, list of group sizes, list of group start positions,
distance between members, step size or prime number to start at. These can be passed as whole number(s) or percentage(s), while method l_starts
also accepts 'auto'
.
Here, we will take a look at the different methods.
greedy
uses group size for dividing up the data.
Greedy refers to each group grabbing as many elements as possible (up to the specified size), meaning that there might be fewer elements available to the last group, but that all other groups than the last are guaranteed to have the size specified.
Example
We have a vector with 57 values. We want to have group sizes of 10.
The greedy splitter will return groups with this many values in them:
10, 10, 10, 10, 10, 7
By setting force_equal = TRUE
, we discard the last group if it contains fewer
values than the other groups.
Example
We have a vector with 57 values. We want to have group sizes of 10.
The greedy splitter with force_equal set to TRUE will return groups with this many values in them:
10, 10, 10, 10, 10
meaning that 7 values have been discarded.
n_dist
uses a specified number of groups to divide up the data.
First, it creates equal groups as large as possible. Then, it distributed
any excess data points across the groups.
Example
We have a vector with 57 values. We want to get back 5 groups.
'n_dist' with default settings would return groups with this many values in them:
11, 11, 12, 11, 12
By setting force_equal = TRUE
, n_dist
will create the largest possible,
equally sized groups by discarding excess data elements.
Example
'n_dist' with force_equal set to TRUE would return groups with this many values in them:
11, 11, 11, 11, 11
meaning that 2 values have been discarded.
n_fill
uses a specified number of groups to divide up the data.
It creates equal groups as large as possible and places any excess data points in the first groups.
By setting descending = TRUE
, it would fill from the the last groups instead.
Example
We have a vector with 57 values. We want to get back 5 groups.
'n_fill' with default settings would return groups with this many values in them:
12, 12, 11, 11, 11
By setting force_equal = TRUE
, n_fill
will create the largest possible,
equally sized groups by discarding excess data elements.
Example
'n_fill' with force_equal set to TRUE would return groups with this many values in them:
11, 11, 11, 11, 11
meaning that 2 values have been discarded.
n_last
uses a specified number of groups to divide up the data.
With default settings, it tries to make the groups as equally sized as possible, but notice that the last group might contain fewer or more elements, if the length of the data is not divisible with the number of groups. All groups, except the last, are guaranteed to contain the same number of elements.
Example
We have a vector with 57 values. We want to get back 5 groups.
'n_last' with default settings would return groups with this many values in them:
11, 11, 11, 11, 13
By setting force_equal = TRUE
, n_last
will create the largest possible,
equally sized groups by discarding excess data elements.
Example
'n_last' with force_equal set to TRUE would return groups with this many values in them:
11, 11, 11, 11, 11
meaning that 2 values have been discarded.
Notice that n_last
will always return the given number of groups. It will never return a
group with zero elements. For some situations that means that the last group will
contain a lot of elements. Asked to divide a vector with 57 elements into 20 groups,
the first 19 groups will contain 2 elements, while the last group will itself contain 19 elements.
Had we instead asked it to divide the vector into 19 groups, we would have had 3 elements
in all groups.
n_rand
uses a specified number of groups to divide up the data.
It creates equal groups as large as possible and places any excess data points randomly in the groups (max. one extra element per group).
Example
We have a vector with 57 values. We want to get back 5 groups.
'n_rand' with default settings could return groups with this many values in them:
12, 11, 11, 11, 12
By setting force_equal = TRUE
, n_rand
will create the largest possible,
equally sized groups by discarding excess data elements.
Example
'n_rand' with force_equal set to TRUE would return groups with this many values in them:
11, 11, 11, 11, 11
meaning that 2 values have been discarded.
l_sizes
divides up the data by a list of group sizes.
Excess data points are placed in extra group at the end.
n
is a list
/vector
of group sizes
Example
We have a vector with 57 values. We want to get back 3 groups containing 20%, 30% and 50% of the data points.
'l_sizes' with n = c(0.2, 0.3) would return groups with this many values in them:
11, 17, 29
By setting force_equal = TRUE
, l_sizes
discard any excess elements.
Example
'l_sizes' with n = c(0.2, 0.3) and force_equal set to TRUE would return groups with this many values in them:
11, 17
meaning that 29 values have been discarded.
l_starts
starts new groups at specified values of vector.
n
is a list
of starting positions. Skip values by c(value, skip_to_number)
where skip_to_number
is the nth appearance of the value in the vector
. Groups automatically start from first data point.
If passing n = 'auto'
the starting positions are automatically found with find_starts()
.
If data is a data frame
, starts_col
must be set to indicate the column to match starts.
Set starts_col
to 'index'
or '.index'
for matching with row names. 'index'
first looks for column named 'index'
in data, while '.index'
completely ignores any potential column in data named '.index'
.
Example
We have a vector with 57 values ranging from (1:57). We want to get back groups starting at specific values in the vector.
'l_starts' with n = c(1, 3, 7, 25, 50) would return groups with this many values in them:
2, 4, 18, 25, 8
force_equal
does not have any effect with method l_starts
.
Groups can start at nth appearance of the value by using c(value, skip_to_number)
.
Example
We have a vector with the values c("a", "e", "o", "a", "e", "o") and want to start groups at the first "a", the first following "e" and the second following "o".
'l_starts' with n = list("a", "e", c("o", 2)) would return groups with this many values in them:
1, 4, 1
Using the find_starts()
function, l_starts
is capable of finding the beginning of groups automatically.
A group start is a value which differs from the previous value.
Example
We have a vector with the values c("a", "a", "o", "o", "o", "a", "a") and want to automatically discover groups of data and group them.
'l_starts' with n = 'auto' would return groups with this many values in them:
2, 3, 2
find_starts()
finds group starts in a given vector.
A group start is a value which differs from the previous value.
Setting return_index = TRUE
returns indices of group starts.
Example
We have a vector with the values c("a", "a", "o", "o", "o", "a", "a") and want to automatically discover group starts.
find_starts() would return these group starts:
"a", "o", "a"
find_missing_starts()
tells you the values and (optionally) skip-to numbers that would be recursively removed when using the l_starts
method with the remove_missing_starts
argument set to TRUE
.
Set return_skip_numbers = FALSE
to get only the missing values without the skip-to numbers.
Example
We have a vector with the values c("a", "a", "o", "o", "o", "a", "a") and a vector of starting positions c("a","d","o","p","a").
find_missing_starts() would return this list of values and skip_to numbers:
list(c("d",1), c("p",1))
every
greedily combines every n
th data point.
Example
We have a vector with 57 values. We want to get back 5 groups.
'every' with default settings would return groups with this many values in them:
12, 12, 11, 11, 11
By setting force_equal = TRUE
, every
discards the last length % n
data points
such that all groups have the same size.
Example
'every' with force_equal set to TRUE would return groups with this many values in them:
11, 11, 11, 11, 11
meaning that 2 values have been discarded.
staircase
uses a step size to divide up the data.
For each group, the group size will be step size multiplied with the group index.
Example
We have a vector with 57 values. We specify a step size of 5.
'staircase' with default settings would return groups with this many values in them:
5, 10, 15, 20, 7
By setting force_equal = TRUE
, staircase
will discard the last group if it does
not contain the expected values (step size multiplied by group index).
Example
'staircase' with force_equal set to TRUE would return groups with this many values in them:
5, 10, 15, 20
meaning that 7 values have been discarded.
When using the staircase
method, the last group might not have the size of the second last group + step size.
Use %staircase%
to find the remainder.
If the last group has the size of the second last group + step size, %staircase%
will return 0
.
Example
%staircase% on a vector with size 57 and step size of 5 would look like this:
57 %staircase% 5
and return:
7
meaning that the last group would contain 7 values
primes
creates groups with sizes of primary numbers in a staircasing design.
n
is the prime number to start at (size of first group).
Prime numbers are generated with the numbers
package by Hans Werner Borchers.
Example
We have a vector with 57 values. We specify n (start at) as 5.
'primes' with default settings would return groups with this many values in them:
5, 7, 11, 13, 17, 4
By setting force_equal = TRUE
, primes
will discard the last group if it does
not contain the expected number of values.
Example
'primes' with force_equal set to TRUE would return groups with this many values in them:
5, 7, 11, 13, 17
meaning that 4 values have been discarded.
When using the primes
method, the last group might not have the size of the associated prime number, if there are not enough elements. Use %primes%
to find the remainder.
Returns 0
if the last group has the size of the associated prime number.
Example
%primes% on a vector with size 57 and n (start at) as 5 would look like this:
57 %primes% 5
and return:
4
meaning that the last group would contain 4 values
There are currently 4 methods for balancing on ID level in balance()
.
Balances on ID level only. It makes sure there are the same number of IDs in each category. This might lead to a different number of rows between categories.
Attempts to level the number of rows per category, while only removing/adding entire IDs.
This is done in 2 steps:
If a category needs to add all its rows one or more times, the data is repeated.
Iteratively, the ID with the number of rows closest to the lacking/excessive number of rows is added/removed. This happens until adding/removing the closest ID would lead to a size further from the target size than the current size. If multiple IDs are closest, one is randomly sampled.
Distributes the lacking/excess rows equally between the IDs. If the number to distribute can not be equally divided, some IDs will have 1 row more/less than the others.
Balances the IDs within their categories, meaning that all IDs in a category will have the same number of rows.
These are the arguments for group_factor()
, group()
, splt()
, fold()
, partition()
Type: data frame
or vector
The data to process.
Used in: group_factor()
, group()
, splt()
, fold()
, partition()
Type: integer
, numeric
, character
, or list
n
represents either number of groups (default), group size, list of group sizes, list of group starts, step size or prime number to start at, depending on which method is specified.
n
can be given as a whole number(s) (n > 1) or as percentage(s) (0 < n < 1).
Method l_starts
allows n = 'auto'
.
Used in: group_factor()
, group()
, splt()
Type: character
Choose which method to use when dividing up the data.
Available methods: greedy
, n_dist
, n_fill
, n_last
, n_rand
, l_starts
, l_sizes
, staircase
, or primes
Used in: group_factor()
, group()
, splt()
, fold()
Type: character
Name of column with values to match in method l_starts
when data
is a data frame
.
Pass 'index'
or '.index'
to use rownames. 'index'
first looks for column named 'index'
in data
, while '.index'
completely ignores any potential column in data named '.index'
.
Used in: group_factor()
, group()
, splt()
Type: logical
(TRUE
or FALSE
)
If you need groups with the exact same size, set force_equal = TRUE
.
Implementation is different in the different methods. Read more in their sections above.
Be aware that this setting discards excess datapoints!
Used in: group_factor()
, group()
, splt()
, partition()
Type: logical
(TRUE
or FALSE
)
If you set n
to 0
, you get an error.
If you don't want this behavior, you can set allow_zero = TRUE
, and (depending on
the function) you will get the following output:
group_factor()
will return the factor with NA
s instead of numbers. It will be
the same length as expected.
group()
will return the expected data frame with NA
s instead of a grouping factor.
splt()
functions will return the given data (data frame
or vector
) in the same list
format as if it had been split.
Used in: group_factor()
, group()
, splt()
Type: logical
(TRUE
or FALSE
)
In methods like n_fill
where it makes sense to change the direction of the method, you can use this argument.
In n_fill
it fills up the excess data points starting from the last group instead of the first.
NB. Only some of the methods can use this argument.
Used in: group_factor()
, group()
, splt()
Type: logical
(TRUE
or FALSE
)
After creating the the grouping factor using the chosen method, it is possible to randomly reorganize it before returning it. Notice that this applies to all the functions that allows for the argument, as group()
and splt()
uses the grouping factor!
Used in: group_factor()
, group()
, splt()
N.B. fold()
and partition()
always uses some randomization.
Type: character
Name of added grouping factor column. Allows multiple grouping factors in a data frame
.
Used in: group()
Type: logical
(TRUE
or FALSE
)
Recursively remove elements from the list of starts that are not found. For method l_starts
only.
Used in: group_factor()
, group()
, splt()
Type: integer
or numeric
k
represents either number of folds (default), fold size, or step size, depending on which method is specified.
k
can be given as a whole number (k > 1) or as a percentage (0 < k < 1).
Used in: fold()
Type: integer
or numeric
Size(s) of partition(s). Passed as vector
if specifying multiple partitions.
p
can be given as whole number(s) (p > 1) or as percentage(s) (0 < p < 1).
Used in: partition()
Type: categorical vector
or factor
(passed as column name)
Categorical variable to balance between the groups.
E.g. when predicting a binary variable ('a' or 'b'), we usually want both classes represented in every fold and partition.
N.B. If also passing id_col
, cat_col
should be a constant within IDs.
E.g. a participant must always have the same diagnosis ('a' or 'b') throughout the dataset.
Otherwise, the participant might be placed in multiple folds.
Used in: fold()
, partition()
Type: numerical vector
(passed as column name)
Numerical variable to balance between groups.
N.B. When used with id_col
, values for each ID are aggregated using id_aggregation_fn
before being balanced.
N.B. When passing num_col
, the method
argument is not used.
Used in: fold()
, partition()
Type: factor
(passed as column name)
factor
with IDs. This will be used to keep all rows with an ID in the same group (if possible).
E.g. If we have measured a participant multiple times and want to see the effect of time, we want to have all observations of this participant in the same fold/partition.
Used in: fold()
, partition()
Type: function
Function for aggregating values in num_col
for each ID, before balancing by num_col
.
N.B. Only used when num_col
and id_col
are both specified.
Used in: fold()
, partition()
Type: integer
or numeric
How many levels of extreme pairing to do when balancing groups by num_col
.
Extreme pairing: Rows/pairs are ordered as smallest, largest, second smallest, second largest, etc. If extreme_pairing_levels > 1
, this is done "recursively" on the extreme pairs.
N.B. Values greater than 1
works best with large datasets. Always check if an increase actually makes the groups more balanced. There are examples of how to do this, and more detailed descriptions of the implementations, in the functions' help files (?fold
and ?partition
).
Used in: fold()
, partition()
Type: integer
or numeric
Number of fold columns to create. This is useful for repeated cross-validation.
If num_fold_cols > 1
, columns will be named ".folds_1"
, ".folds_2"
, etc.
Otherwise simply ".folds"
.
N.B. If unique_fold_cols_only
is TRUE
, we can end up with fewer columns than specified, see max_iters
.
N.B. If data
has existing fold columns, see handle_existing_fold_cols
.
Used in: fold()
Type: logical
(TRUE
or FALSE
)
Check if the fold columns are identical and keep only the unique columns.
N.B. As the number of column comparisons can be time consuming,
we can run this part in parallel. See parallel
.
N.B. We can end up with fewer columns than specified in
num_fold_cols
, see max_iters
.
N.B. Only used when num_fold_cols > 1
or data
has existing fold columns.
Used in: fold()
Type: logical
(TRUE
or FALSE
)
Maximum number of attempts at reaching
num_fold_cols
unique fold columns.
When only keeping the unique fold columns, we risk having fewer columns than expected.
Hence, we repeatedly create the missing columns and remove those that are not unique.
This is done until we have num_fold_cols
unique fold columns,
or we have attempted max_iters
times.
In some cases, it is not possible to create num_fold_cols
unique combinations of the dataset, e.g.
when specifying cat_col
, id_col
and num_col
.
max_iters
specifies when to stop trying.
N.B. We can end up with fewer columns than specified in num_fold_cols
.
N.B. Only used num_fold_cols > 1
.
Used in: fold()
Type: character
How to handle existing fold columns.
Either "keep_warn"
, "keep"
, or "remove"
.
To add extra fold columns, use "keep"
or "keep_warn"
.
Note that existing fold columns might be renamed.
To replace the existing fold columns, use "remove"
.
Used in: fold()
Type: logical
(TRUE
or FALSE
)
Whether to parallelize the fold column comparisons,
when unique_fold_cols_only
is TRUE
.
N.B. Requires a registered parallel backend. Like with doParallel:registerDoParallel
.
Used in: fold()
Type: logical
(TRUE
or FALSE
)
Return list
of partitions (TRUE
) or a grouped data frame
(FALSE
).
Used in: partition()
These are the arguments for balance()
, upsample()
, downsample()
Type: data frame
The data to process.
Used in: balance()
, upsample()
, downsample()
Type: character
, numeric
Size to balance categories to.
Either a specific number, given as a whole number, or one of the following strings: "min"
, "max"
, "mean"
, "median"
.
Used in: balance()
, upsample()
, downsample()
Type: categorical vector
or factor
(passed as column name)
Categorical variable to balance sample size by.
Used in: balance()
, upsample()
, downsample()
Type: factor
(passed as column name)
factor
with IDs. IDs are considered entities, e.g. allowing us to add or remove all rows for an ID. How this is used is up to the id_method
.
E.g. If we have measured a participant multiple times and want make sure that we keep all these measurements together. Then we would either remove/add all measurements for the participant or leave in all measurements for the participant.
Used in: balance()
, upsample()
, downsample()
Type: character
Method for balancing the IDs.
Available ID methods are: n_ids
, n_rows_c
, or nested
.
Used in: balance()
, upsample()
, downsample()
Type: logical
Whether to add a column with 1
s for added rows, and 0
s for original rows.
Used in: balance()
, upsample()
, downsample()
Type: character
Name of column marking new rows.
Used in: balance()
, upsample()
, downsample()
We will be using the method n_dist
on a data frame
to showcase the functions. Afterwards we will use and compare the methods.
Note, that you can also use vector
s as data with all the functions.
See the necessary attached packages for running the examples under [Attach Packages].
data frame
df <- data.frame( "x" = c(1:12), "species" = factor(rep(c('cat', 'pig', 'human'), 4)), "age" = sample(c(1:100), 12) )
group_factor()
groups <- group_factor(df, 5, method = 'n_dist') groups df$groups <- groups df %>% kable(align = 'c')
df %>% group_by(groups) %>% summarize(mean_age = mean(age)) %>% kable(align = 'c')
Getting an equal number of elements per group with group_factor()
.
Notice that we discard the excess values so all groups contain the same amount of elements. Since the grouping factor is shorter than the data frame, we can't combine them as they are. A way to do so would be to shorten the data frame to be the same length as the grouping factor.
data frame
df <- data.frame( "x" = c(1:12), "species" = factor(rep(c('cat', 'pig', 'human'), 4)), "age" = sample(c(1:100), 12) )
group_factor()
with force_equal
groups <- group_factor(df, 5, method = 'n_dist', force_equal = TRUE) groups plyr::count(groups) %>% rename(group = x, size = freq) %>% kable(align = 'c')
data frame
and grouping factor First we make the data frame
the same size as the grouping factor. Then we add the grouping factor to the data frame
.
df <- head(df, length(groups)) %>% mutate(group = groups) df %>% kable(align = 'c')
data frame
df <- data.frame( "x" = c(1:12), "species" = factor(rep(c('cat', 'pig', 'human'), 4)), "age" = sample(c(1:100), 12) )
group()
df_grouped <- group(df, 5, method = 'n_dist') df_grouped %>% kable(align = 'c')
2.2 Using group()
in magrittr
pipeline to get mean age
df_means <- df %>% group(5, method = 'n_dist') %>% summarise(mean_age = mean(age)) df_means %>% kable(align = 'c')
Getting an equal number of elements per group with group()
.
Notice that we discard the excess rows/elements so all groups contain the same amount of elements.
data frame
df <- data.frame( "x" = c(1:12), "species" = factor(rep(c('cat', 'pig', 'human'), 4)), "age" = sample(c(1:100), 12) )
group()
with force_equal
df_grouped <- df %>% group(5, method = 'n_dist', force_equal = TRUE) df_grouped %>% kable(align = 'c')
data frame
df <- data.frame( "x" = c(1:12), "species" = factor(rep(c('cat', 'pig', 'human'), 4)), "age" = sample(c(1:100), 12) )
splt()
df_list <- splt(df, 5, method = 'n_dist') df_list %>% kable(align = 'c')
splt()
without using kable()
to visualize it.splt()
uses base::split()
to split the data by the grouping factor.v <- c(1:6) splt(v, 3, method = 'n_dist')
Getting an equal number of elements per group with splt()
.
Notice that we discard the excess rows/elements so all groups contain the same amount of elements.
data frame
df <- data.frame( "x" = c(1:12), "species" = factor(rep(c('cat', 'pig', 'human'), 4)), "age" = sample(c(1:100), 12) )
splt()
with force_equal
df_list <- splt(df, 5, method = 'n_dist', force_equal = TRUE) df_list %>% kable(align = 'c')
data frame
df <- data.frame( "participant" = factor(rep(c('1', '2', '3', '4', '5', '6'), 3)), "age" = rep(sample(c(1:100), 6), 3), "diagnosis" = factor(rep(c('a', 'b', 'a', 'a', 'b', 'b'), 3)), "score" = sample(c(1:100), 3 * 6) ) df <- df %>% arrange(participant) # Remove index rownames(df) <- NULL # Add session info df$session <- rep(c('1','2', '3'), 6) kable(df, align = 'c')
fold()
without balancing df_folded <- fold(df, 3, method = 'n_dist') # Order by folds df_folded <- df_folded %>% arrange(.folds) kable(df_folded, align = 'c')
fold()
with balancing but without id_col df_folded <- fold(df, 3, cat_col = 'diagnosis', method = 'n_dist') # Order by folds df_folded <- df_folded %>% arrange(.folds) kable(df_folded, align = 'c')
Let's count how many of each diagnosis there are in each group.
df_folded %>% count(.folds, diagnosis) %>% kable(align='c')
fold()
with id_col
but without balancing df_folded <- fold(df, 3, id_col = 'participant', method = 'n_dist') # Order by folds df_folded <- df_folded %>% arrange(.folds) # Remove index (Looks prettier in the table!) rownames(df_folded) <- NULL kable(df_folded, align = 'c')
Let's see how participants were distributed in the groups.
df_folded %>% count(.folds, participant) %>% kable(align='c')
fold()
with balancing and with id_col
fold()
first divides up the data frame by cat_col
and then create n
folds for both diagnoses. As there are only 3 participants per diagnosis, we can maximally create 3 folds in this scenario.
df_folded <- fold(df, 3, cat_col = 'diagnosis', id_col = 'participant', method = 'n_dist') # Order by folds df_folded <- df_folded %>% arrange(.folds) kable(df_folded, align = 'c')
Let's count how many of each diagnosis there are in each group and find which participants are in which groups.
df_folded %>% count(.folds, diagnosis, participant) %>% kable(align='c')
data frame
df <- data.frame( "participant" = factor(rep(c('1', '2', '3', '4', '5', '6'), 3)), "age" = rep(sample(c(1:100), 6), 3), "diagnosis" = factor(rep(c('a', 'b', 'a', 'a', 'b', 'b'), 3)), "score" = sample(c(1:100), 3 * 6) ) df <- df %>% arrange(participant) # Remove index rownames(df) <- NULL # Add session info df$session <- rep(c('1','2', '3'), 6) kable(df, align = 'c')
partition()
without balancing df_partitioned <- partition(df, 0.3, list_out = FALSE) # Order by partitions df_partitioned <- df_partitioned %>% arrange(.partitions) # Partition Sizes df_partitioned %>% count(.partitions) %>% kable(align = 'c') kable(df_partitioned, align = 'c')
partition()
with balancing but without id_col
df_partitioned <- partition(df, 0.3, cat_col = 'diagnosis', list_out = FALSE) # Order by partitions df_partitioned <- df_partitioned %>% arrange(.partitions) kable(df_partitioned, align = 'c')
Let's count how many of each diagnosis there are in each partition.
df_partitioned %>% count(.partitions, diagnosis) %>% kable(align='c')
partition()
with id_col
but without balancing df_partitioned <- partition(df, 0.5, id_col = 'participant', list_out = FALSE) # Order by partitions df_partitioned <- df_partitioned %>% arrange(.partitions) kable(df_partitioned, align = 'c')
Let's see how participants were distributed in the partitions.
df_partitioned %>% count(.partitions, participant) %>% kable(align='c')
partition()
with cat_col
and with id_col
partition()
first divides up the data frame by cat_col
and then create n
partitions for both diagnoses.
df_partitioned <- partition( data = df, p = 0.5, cat_col = 'diagnosis', id_col = 'participant', list_out = FALSE ) # Order by folds df_partitioned <- df_partitioned %>% arrange(.partitions) kable(df_partitioned, align = 'c')
Let's count how many of each diagnosis there are in each group and find which participants are in which partitions.
df_partitioned %>% count(.partitions, diagnosis, participant) %>% kable(align='c')
dataframe
set.seed(2)
df <- data.frame( "participant" = factor(rep(c('1', '2', '3', '4', '5', '6'), 3)), "age" = rep(sample(c(1:100), 6), 3), "diagnosis" = factor(rep(c('a', 'b', 'a', 'a', 'b', 'b'), 3)), "score" = sample(c(1:100), 3 * 6) ) df <- df %>% arrange(participant) # Add session info df$session <- rep(c('1','2', '3'), 6) # Sample dataset to get imbalances df <- df %>% sample_frac(0.7) %>% arrange(participant) # Remove index rownames(df) <- NULL # Counts df %>% count(diagnosis, participant) %>% kable(align = 'c') df %>% count(diagnosis) %>% kable(align = 'c') kable(df, align = 'c')
balance()
for downsampling without id_col
df_balanced <- balance(df, "min", cat_col = "diagnosis") %>% arrange(diagnosis, participant) # Counts df_balanced %>% count(diagnosis) %>% kable(align = 'c') kable(df_balanced, align = 'c')
balance()
for downsampling with id_col
and id_method
n_rows_c
df_balanced <- balance(df, "min", cat_col = "diagnosis", id_col = "participant", id_method = "n_rows_c") %>% arrange(diagnosis, participant) # Partition Sizes df_balanced %>% count(diagnosis) %>% kable(align = 'c') kable(df_balanced, align = 'c')
Let's count how many of each participant there are in each diagnosis.
df_balanced %>% count(diagnosis, participant) %>% kable(align='c')
balance()
for upsampling without id_col
df_balanced <- balance(df, "max", cat_col = "diagnosis") %>% arrange(diagnosis, participant) # Counts df_balanced %>% count(diagnosis) %>% kable(align = 'c') kable(df_balanced, align = 'c')
balance()
for upsampling with id_col
and id_method
n_rows_c
df_balanced <- balance(df, "max", cat_col = "diagnosis", id_col = "participant", id_method = "n_rows_c") %>% arrange(diagnosis, participant) # Partition Sizes df_balanced %>% count(diagnosis) %>% kable(align = 'c') kable(df_balanced, align = 'c')
Let's count how many of each participant there are in each diagnosis.
df_balanced %>% count(diagnosis, participant) %>% kable(align='c')
balance()
for downsampling to a specific size without id_col
df_balanced <- balance(df, 3, cat_col = "diagnosis") %>% arrange(diagnosis, participant) # Counts df_balanced %>% count(diagnosis) %>% kable(align = 'c') kable(df_balanced, align = 'c')
Let's count how many of each participant there are in each diagnosis.
df_balanced %>% count(diagnosis, participant) %>% kable(align='c')
balance()
for downsampling to a specific size with id_col
and id_method
n_rows_c
df_balanced <- balance(df, 3, cat_col = "diagnosis", id_col = "participant", id_method = "n_rows_c") %>% arrange(diagnosis, participant) # Partition Sizes df_balanced %>% count(diagnosis) %>% kable(align = 'c') kable(df_balanced, align = 'c')
Let's count how many of each participant there are in each diagnosis.
df_balanced %>% count(diagnosis, participant) %>% kable(align='c')
data frame
df <- data.frame( "x" = c(1:12), "species" = factor(rep(c('cat', 'pig', 'human'), 4)), "age" = sample(c(1:100), 12) )
group_factor()
with randomize
set to TRUE
groups <- group_factor(df, 5, method = 'n_dist', randomize = TRUE) groups
splt()
with randomize
set to TRUE
df_list <- splt(df, 5, method = 'n_dist', randomize = TRUE) df_list %>% kable(align = 'c')
In this section, we will take a look at the outputs we get from the different methods.
Below you'll see a data frame with counts of group elements when dividing up the same data with the different n_*
methods.
The forced-equal column is simply with the force_equal
set to TRUE.
forced_equal: Since this is a setting to make sure that all groups are of the same size, it makes sense that all the groups have the same size.
n_dist: compared to forced_equal we see the 3 datapoints that forced_equal had discarded. These are distributed across the groups (in this example group 2,4 and 6)
n_fill: The 3 extra datapoints are located at the first 3 groups. Had we set descending = TRUE
,
it would have been the last 3 groups instead.
n_last: We see that n_last
creates equal group sizes in all but the last group. This means that the last group can sometimes have a group size, which is very large or small compared to the other groups. Here it is a third larger than the other groups.
n_rand: The extra datapoints are placed randomly and so we would see the extra datapoints located at different groups if we ran the script again. Unless we use set.seed()
before running the function.
# # Examples to show difference between methods # This could be made interactive! This way you could test what happens in different situations by # by simply moving a slider! # vec <- c(1:57) n <- 6 if (exists ('n_meth_v57n6')){ rm(n_meth_v57n6) } for (meth in c('n_dist', 'n_fill' ,'n_last','n_rand')){ data_temp <- data.frame(plyr::count(group_factor(vec, n, method = meth))) names(data_temp)[names(data_temp)=="freq"] <- meth if (exists ('n_meth_v57n6')) { n_meth_v57n6 <- cbind(n_meth_v57n6, data_temp) } else { n_meth_v57n6 <- data_temp } } forced_equal <- plyr::count(group_factor(vec, n, method = 'n_last', force_equal = TRUE)) n_meth_v57n6$forced_equal <- forced_equal$freq n_meth_v57n6 <- n_meth_v57n6[ , !duplicated(colnames(n_meth_v57n6))] # gather() data frame for plotting data_plot <- n_meth_v57n6 %>% gather(method, group_size,-1) upper_limit <- max(data_plot$group_size)+1 lower_limit <- min(data_plot$group_size)-1 v57n6_plot <- ggplot(data_plot, aes(x, group_size)) ## Output # Data frame n_meth_v57n6 # Plot v57n6_plot + geom_point() + scale_y_continuous(limit = c(lower_limit, upper_limit), breaks = round(seq(lower_limit, upper_limit, by = 2),1)) + #scale_y_continuous(limit = c(lower_limit, upper_limit))+ facet_wrap('method', ncol=1) + labs(x = 'group', y = 'group Size', title = 'Distribution of Elements in groups')+ theme_bw()+ theme(axis.text.y = element_text(size=9), axis.text.x = element_text(size=9))
Here is another example.
vec <- c(1:117) n <- 11 if (exists ('n_meth_v117n11')){ rm(n_meth_v117n11) } for (meth in c('n_dist', 'n_fill' ,'n_last','n_rand')){ data_temp <- data.frame(plyr::count(group_factor(vec, n, method = meth))) names(data_temp)[names(data_temp)=="freq"] <- meth if (exists ('n_meth_v117n11')) { n_meth_v117n11 <- cbind(n_meth_v117n11, data_temp) } else { n_meth_v117n11 <- data_temp } } forced_equal <- plyr::count(group_factor(vec, n, method = 'n_last', force_equal = TRUE)) n_meth_v117n11$forced_equal <- forced_equal$freq n_meth_v117n11 <- n_meth_v117n11[ , !duplicated(colnames(n_meth_v117n11))] # gather() data frame for plotting data_plot <- n_meth_v117n11 %>% gather(method, group_size,-1) v117n11_plot <- ggplot(data_plot, aes(x, group_size)) upper_limit <- max(data_plot$group_size)+1 lower_limit <- min(data_plot$group_size)-1 ## Output # Data frame n_meth_v117n11 # Plot v117n11_plot + geom_point() + scale_y_continuous(limit = c(lower_limit, upper_limit), breaks = round(seq(lower_limit, upper_limit, by = 2),1)) + facet_wrap('method', ncol=1) + labs(x = 'group', y = 'group Size', title = 'Distribution of Elements in groups')+ theme_bw()+ theme(axis.text.y = element_text(size=9), axis.text.x = element_text(size=9))
Below you will see group sizes when using the method greedy
and asking for group sizes of 8, 15, 20. What should become clear is that only the last group can have a different group size than what we asked for. This is important if, say, you want to split a time series into groups of 100 elements, but the time series is not divisible with 100. Then you could use force_equal
to remove the excess elements, if you need equal groups.
With a size of 8, we get 13 groups. The last group (13) only contains 4 elements, but all the other groups contain 8 elements as specified.
With a size of 15, we get 7 groups. The last group (7) contains only 10 elements, but all the other groups contain 15 elements as specified.
With a size of 20, we get 5 groups. As 20 is divisible with the 100 elements that the split vector contained, the last group also contains 20 elements, and so we have equal groups.
vec <- c(1:100) if (exists ('greedy_data')){ rm(greedy_data) } for (n in c(8,15,20)){ group_sizes <- plyr::count(group_factor(vec, n, method='greedy')) data_temp <- data.frame(group_sizes, 'Size' = factor(n)) if (exists ('greedy_data')) { greedy_data <- rbind(greedy_data, data_temp) } else { greedy_data <- data_temp } } greedy_plot <- ggplot(greedy_data, aes(x, freq, color=Size)) greedy_plot + geom_point() + labs(x = 'group', y = 'group Size', title = 'Greedy Distribution of Elements in groups', color = 'Size') + theme_bw()+ theme(plot.margin = unit(c(1,1,1,1), "cm"))+ theme(axis.text.y = element_text(size=9), axis.text.x = element_text(size=9))
Below you'll see a plot with the group sizes at each group when using step sizes 2, 5, and 11.
At a step size of 2 elements it simply increases 2 for each group, until the last group (32) where it runs out of elements. Had we set force_equal = TRUE
, this last group would have been discarded, because of the lack of elements.
At a step size of 5 elements it increases with 5 every time. Because of this it runs out of elements faster. Again we see that the last group (20) has fewer elements.
At a step size of 11 elements it increases with 11 every time. It seems that the last group is not too small, but it can be hard to see on this scale. Actually, the last group misses 1 element to be complete and so would have been discarded if force_equal
was set to TRUE
.
vec <- c(1:1000) if (exists ('staircase_data')){ rm(staircase_data) } for (n in c(2, 5, 11)){ group_sizes <- plyr::count(group_factor(vec, n, method='staircase')) data_temp <- data.frame(group_sizes, 'step_size' = factor(n)) if (exists ('staircase_data')) { staircase_data <- rbind(staircase_data, data_temp) } else { staircase_data <- data_temp } } staircase_plot <- ggplot(staircase_data, aes(x, freq, color=step_size)) staircase_plot + geom_point() + #scale_x_continuous(breaks = round(seq(1, max(data_temp$x), by = 2),1))+ labs(x = 'group', y = 'group Size', title = 'Staircasing Distribution of Elements in groups', color = 'Step Size') + theme_bw()+ theme(axis.text.y = element_text(size=9), axis.text.x = element_text(size=7))
Below we will take a quick look at the cumulative sum of group elements to get an idea of what is going on under the hood.
Remember that the split vector had 1000 elements? That is why they all stop at 1000 on the y-axis. There are simply no more elements left!
staircase_data <- staircase_data %>% group_by(step_size) %>% mutate(cumsum = cumsum(freq)) staircase_cumulative_plot <- ggplot(staircase_data, aes(x, cumsum, color=step_size)) staircase_cumulative_plot + geom_point() + labs(x = 'group', y = 'Cumulative sum of group sizes', title = 'Staircasing Cumulative Sum of group Sizes', color = 'Step Size') + theme_bw()+ theme(axis.text.y = element_text(size=9), axis.text.x = element_text(size=7))
Below you'll see a plot with the group sizes at each group when starting from prime numbers 2, 5, and 11.
vec <- c(1:1000) if (exists ('primes_data')){ rm(primes_data) } for (n in c(2, 5, 11)){ group_sizes <- plyr::count(group_factor(vec, n, method='primes')) data_temp <- data.frame(group_sizes, 'start_at' = factor(n)) if (exists ('primes_data')) { primes_data <- rbind(primes_data, data_temp) } else { primes_data <- data_temp } } primes_plot <- ggplot(primes_data, aes(x, freq, color=start_at)) primes_plot + geom_point() + #scale_x_continuous(breaks = round(seq(1, max(data_temp$x), by = 2),1))+ labs(x = 'group', y = 'group Size', title = 'Prime numbers method - Elements per groups', color = 'Start at') + theme_bw()+ theme(axis.text.y = element_text(size=9), axis.text.x = element_text(size=7))
Below we will take a quick look at the cumulative sum of group elements to get an idea of what is going on under the hood.
Because the split vector had 1000 elements, it stops at 1000 on the y-axis. There are simply no more elements left!
primes_data <- primes_data %>% group_by(start_at) %>% mutate(cumsum = cumsum(freq)) primes_cumulative_plot <- ggplot(primes_data, aes(x, cumsum, color=start_at)) primes_cumulative_plot + geom_point() + labs(x = 'group', y = 'Cumulative sum of group sizes', title = 'Primes Cumulative Sum of group Sizes', color = 'Start At') + theme_bw()+ theme(axis.text.y = element_text(size=9), axis.text.x = element_text(size=7))
You have reached the end! Now celebrate by taking the week off, splitting data and laughing!
Any scripts or data that you put into this service are public.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.