cyclic_encoding: Cyclic encoding of date-times

In lubridate: Make Dealing with Dates a Little Easier

Cyclic encoding of date-times

Description

Encode a date-time object into a cyclic coordinate system in which the distances between two pairs of dates separated by the same time duration are the same.

Usage

cyclic_encoding(
  x,
  periods,
  encoders = c("sin", "cos"),
  week_start = getOption("lubridate.week.start", 7)
)

Arguments

x	a date-time object
periods	a character vector of periods. Follows same specification as period and floor_date functions.
encoders	names of functions to produce the encoding. Defaults to "sin" and "cos". Names of any predefined functions accepting a numeric input are allowed.
week_start	week start day (Default is 7, Sunday. Set lubridate.week.start to override). Full or abbreviated names of the days of the week can be in English or as provided by the current locale.

Details

Machine learning models don't know that December 31st and January 1st are close in our human calendar sense. cyclic_encoding makes it obvious to the machine learner that two calendar dates are close by mapping the dates onto the circle.

Value

a numeric matrix with number of columns equal length(periods) * length(types).

Examples

times <- ymd_hms("2019-01-01 00:00:00") + hours(0:23)
cyclic_encoding(times, c("day", "week", "month"))
plot(cyclic_encoding(times, "1d"))
plot(cyclic_encoding(times, "2d"), xlim = c(-1, 1))
plot(cyclic_encoding(times, "4d"), xlim = c(-1, 1))

lubridate documentation built on Sept. 27, 2023, 5:07 p.m.