Creates data with identical statistics (metamers) using an iterative
algorithm proposed by Matejka & Fitzmaurice (2017)
Implements the algorithm proposed by Matejka & Fitzmaurice (2017) to create metamers (datasets with identical statistical properties but very different graphs).
In colour theory, metamers are colours that have very different wavelength distribution but are perceived as equal by out visual system. This happens because out eyes essentially summarise a continuous distribution of wavelength by just 3 numbers: the amount that each type of cone cell is exited. Colour metamerism is how artists can reproduce so many colours with a few pigments, or how PC monitors use only 3 lights to show colourful pictures.
Statistical transformations such as mean, standard deviation and correlation behave very similarly in that they summarise data with just a few numbers for the benefit of our limited cognitive capacity. Thus, statistical metamers are sets of data that share some statistical properties.
You can install the development version of metamer with:
You can construct metamers from a starting dataset and a vector of statistical properties to remain constant (by default, up to 2 significant figures).
library(metamer)# Start with the datasaurus# install.packages("datasauRus")start <- subset(datasauRus::datasaurus_dozen, dataset == "dino")start$dataset <- NULL# And we want to preserve means and correlationmean_cor <- delayed_with(mean(x), mean(y), cor(x, y))N <- 20000set.seed(42) # To make results reproduciblemetamers <- metamerize(start, preserve = mean_cor, N = N)print(metamers)#> List of 12782 metamers
We found 12782 metamers. Let's see the final one, with the starting dataset as background.
library(ggplot2)ggplot(metamers[[length(metamers)]], aes(x, y)) +geom_point(data = start, color = "red", alpha = 0.5, size = 0.4) +geom_point()
We can check that the statistical properties have been preserved up to 2 significant figures:
cbind(dino = signif(mean_cor(start), 2),last = signif(mean_cor(metamers[[length(metamers)]]), 2))#> dino last#> [1,] 54.000 54.000#> [2,] 48.000 48.000#> [3,] -0.064 -0.064
However, a semi random cloud of points is not that interesting, so we can specify a minimizing function so that the result is similar to another dataset.
metamerize will start from the last metamer of the previous run if the
data argument is a list of metamers and append the result.
target1 <- subset(datasauRus::datasaurus_dozen, dataset == "x_shape")target1$dataset <- NULL
metamers <- metamerize(metamers,minimize = mean_dist_to(target1),N = N)
Now the result is a bit more impressive.
ggplot(metamers[[length(metamers)]], aes(x, y)) +geom_point(data = start, color = "red", alpha = 0.5, size = 0.4) +geom_point()
We can animate the whole thing. Since 19030 metamers is overkill, first we keep only 200 of them.
library(gganimate)metamers_anim <- trim(metamers, 30*2)ggplot(as.data.frame(metamers_anim), aes(x, y)) +geom_point() +transition_manual(.metamer)#> nframes and fps adjusted to match transition
You can freehand your own starting or target data with the
draw_data() utility, that will open a shiny interface. You might need to install
Metamerizing operations can be chained while changing the minimizing function.
library(magrittr)target2 <- subset(datasauRus::datasaurus_dozen, dataset == "star")target2$dataset <- NULLmetamers <- metamerize(start,preserve = mean_cor,minimize = mean_dist_to(target1),N = N) %>%set_minimize(mean_dist_to(target2)) %>%metamerize(N = N) %>%set_minimize(mean_dist_to(start)) %>%metamerize(N = N)
And the full sequence
trim(metamers, 30*3) %>%as.data.frame() %>%ggplot(aes(x, y)) +geom_point() +transition_manual(.metamer)
NEWS.mdfile to track changes to the package.