tbl_cube: A data cube tbl

Description

A cube tbl stores data in a compact array format where dimension names are not needlessly repeated. They are particularly appropriate for experimental data where all combinations of factors are tried (e.g. complete factorial designs), or for storing the result of aggregations. Compared to data frames, they will occupy much less memory when variables are crossed, not nested.

Usage

tbl_cube(dimensions, measures)

Arguments

dimensions

A named list of vectors. A dimension is a variable whose values are known before the experiment is conducted; they are fixed by design (in reshape2 they are known as id variables). tbl_cubes are dense which means that almost every combination of the dimensions should have associated measurements: missing values require an explicit NA, so if the variables are nested, not crossed, the majority of the data structure will be empty. Dimensions are typically, but not always, categorical variables.

measures

A named list of arrays. A measure is something that is actually measured, and is not known in advance. The dimension of each array should be the same as the length of the dimensions. Measures are typically, but not always, continuous values.

Implementation

Manipulation functions:

select() (M)
summarise() (M), corresponds to roll-up, but rather more limited since there are no hierarchies.
filter() (D), corresponds to slice/dice.
mutate() (M) is not implemented, but should be relatively straightforward given the implementation of summarise.
arrange() (D?) Not implemented: not obvious how much sense it would make

Joins: not implemented. See vignettes/joins.graffle for ideas. Probably straightforward if you get the indexes right, and that's probably some straightforward array/tensor operation.

Details

tbl_cube support is currently experimental and little performance optimisation has been done, but you may find them useful if your data already comes in this form, or you struggle with the memory overhead of the sparse/crossed of data frames. There is no support for hierarchical indices (although I think that would be a relatively straightforward extension to storing data frames for indices rather than vectors).

Examples

Run this code

# NOT RUN {
# The built in nasa dataset records meterological data (temperature,
# cloud cover, ozone etc) for a 4d spatio-temporal dataset (lat, long,
# month and year)
nasa
head(as.data.frame(nasa))

titanic <- as.tbl_cube(Titanic)
head(as.data.frame(titanic))

admit <- as.tbl_cube(UCBAdmissions)
head(as.data.frame(admit))

as.tbl_cube(esoph, dim_names = 1:3)

# Some manipulation examples with the NASA dataset --------------------------

# select() operates only on measures: it doesn't affect dimensions in any way
select(nasa, cloudhigh:cloudmid)
select(nasa, matches("temp"))

# filter() operates only on dimensions
filter(nasa, lat > 0, year == 2000)
# Each component can only refer to one dimensions, ensuring that you always
# create a rectangular subset
# }
# NOT RUN {
filter(nasa, lat > long)
# }
# NOT RUN {
# Arrange is meaningless for tbl_cubes

by_loc <- group_by(nasa, lat, long)
summarise(by_loc, pressure = max(pressure), temp = mean(temperature))
# }