Facet specification for the Grammar of Graphics "add" operator

[paleolimbot]

@clauswilke recently suggested that it would be nice to be able to generate this without resorting to composing plots:

library(ggplot2)
library(patchwork)

p1 <- ggplot(mtcars, aes(disp, mpg)) + geom_point()
p2 <- ggplot(mtcars, aes(wt, mpg)) + geom_point()
p1 + p2

^{Created on 2019-06-06 by the reprex package (v0.2.1)}

I wonder if nested tibbles could be used (currently they fail epicly):

library(ggplot2)
library(tibble)

ggplot(mtcars, aes(x = tibble(disp, wt), y = mpg)) +
  geom_point()
#> Don't know how to automatically pick scale for object of type tbl_df/tbl/data.frame. Defaulting to continuous.
#> Error: Aesthetics must be either length 1 or the same as the data (32): x

^{Created on 2019-06-06 by the reprex package (v0.2.1)}

[thomasp85]

I think this would break every assumption about how plots are declared... at the very least this should be some sort of facet function, but even that is a stretch IMO as the different facets would have different mapping...

what is so wrong with using plot composition for this?

[clauswilke]

Just a quick note from my side. I'm currently traveling and this deserves more careful treatment.

Most importantly, I think the issue has a misleading name. It proposes a specific API, and that's way too early. My point is the following: I was at the SDSS 2019 conference last week and there was a session on the grammar of graphics, and there I learned that the original grammar had a more general approach to laying out plots, and our faceting is a simplified version of it. The original grammar has three fundamental operators, "cross", "add", and "nest", and the example above is an example of "add". I think (need to check more carefully) that "nest" is similar to facet_wrap(), and "cross" is similar to facet_grid(). We have no equivalent to "add", and we cannot combine the operators into a more complex layout.

A related topic is the concept of repeating plots. This is something that Vega has. You can define a generic plot and then repeat it, substituting a different data column for a specified aesthetic for every repeat. An API for this could be something like:

ggplot(mtcars, aes(x = repeat_x, y = mpg)) + 
  geom_point() +
  repeat(repeat_x = list(disp, wt))

But again, the devil is in the detail, and this requires careful thinking through the various use cases, and maybe also looking into how Vega and other systems (e.g., Tableau) do this.

[paleolimbot]

Sorry @clauswilke, I was also traveling and created the issue quickly. I also spent some time trying to get the tibble() specification to work and that's definitely a no go. Claus' syntax works quite well though as a thin wrapper around patchwork:

library(ggplot2)
library(patchwork)
library(rlang)

replace_name <- function(x, name, replace) {
  if (identical(x, name)) {
    return(replace)
  } else if (is.call(x)) {
    return(as.call(lapply(x, replace_name, name, replace)))
  }
  
  x
}

replace_name_quo <- function(q, name, replace) {
  set_expr(q, replace_name(get_expr(q), name, replace))
}

replace_name_aes <- function(mapping, name, replace) {
  for (aesthetic in names(mapping)) {
    mapping[[aesthetic]] <- replace_name_quo(mapping[[aesthetic]], name, replace)
  }
  
  mapping
}

replace_name_plot <- function(plot, name, replace) {
  plot$mapping <- replace_name_aes(plot$mapping, name, replace)
  plot$labels <- ggplot2:::make_labels(plot$mapping)
  plot$layers <- lapply(plot$layers, function(layer) {
    layer$mapping <- replace_name_aes(layer$mapping, name, replace)
    layer
  })
  plot
}

plot_repeat <- function(...) {
  args <- list(...)
  structure(list(repeats = args[1], patchwork_args = args[-1]), class = "plot_repeat")
}

ggplot_add.plot_repeat <- function(object, plot, objectname) {
  name <- sym(names(object$repeats))
  plots <- lapply(object$repeats[[1]], function(q) {
    replace_name_plot(plot, name, get_expr(q))
  })
  
  exec(patchwork::wrap_plots, plots, !!!object$patchwork_args)
}


ggplot(mtcars, aes(x = repeat_x, y = mpg)) + 
  geom_point() +
  plot_repeat(repeat_x = quos(disp, wt), ncol = 1)

^{Created on 2019-06-07 by the reprex package (v0.2.1)}

[clauswilke]

This is nice. Can you use it to make a scatterplot matrix (i.e., repeat both x and y aesthetics)?

To make this fully useful, though, we’d probably want scales to be shared across plots (e.g., if there is a color mapping, we need only one legend), and we’d also need a way to specify position scales for specific panels only. For the latter problem, I’m thinking that scales could take an argument that somehow specifies the panel. (So if you have multiple panels you need to specify a separate scale for each.) I’m just not sure yet how exactly that would look.

[thomasp85]

I really feels like all this should be collected in a new facet function that doesn’t touch the current api of scales etc...

[thomasp85]

To be clear, I think it is very important to acknowledge that ggplot2 is not an implementation of Wilkinson’s grammar of graphics but an adaption with specific deviations to make it more amenable to a consistent programming API. I think that it is very important not to loose that of sight when we consider adding these sort of things.

This is not to detract from the usefulness of such a feature — it sure would be nice — but not at the cost of major complications in API or implementation

[clauswilke]

I'd be happy with putting this into a faceting function. It may be possible to write a new faceting type that can handle all the operators from the original grammar of graphics. I'll look into that when I'm back home and can look at the book.

The issue of specifying scales for specific panels is orthogonal, and it's an existing problem with the current faceting code. We cannot currently set ranges, breaks, or labels by facet. Finding a solution to this problem would be desirable.

[hadley]

TBH I think this is outside the scope of ggplot2. IMO a better integration of positioning across scales and facetting will require a large rewrite.

[clauswilke]

@hadley After reading the relevant parts of the Grammar of Graphics, I agree. Wilkinson approaches positioning quite differently to how things currently work in ggplot2. I'd still like to see a repeat operation, but that could be implemented on top of patchwork just like @paleolimbot demonstrated.

[japhir]

The issue of specifying scales for specific panels is orthogonal, and it's an existing problem with the current faceting code. We cannot currently set ranges, breaks, or labels by facet. Finding a solution to this problem would be desirable.

This is the main reason I currently use patchwork at all. I have some plots where I plot different columns as facets with their own y-axes with a shared x-axis, but because of conventions, some of the "facets" have reversed axes (for good reasons) and others require specific ylims. One also requires a secondary y-axis. This means that presently I create all the plots separately, then using patchwork combine them with the following.

(p1 / p2 / p3 & theme(axis.text.x = element_blank(), legend.pos = "none")) / 
    (p4 & theme(legend.pos = "right"))

But I'm not sure if this should be a separate issue, or is beyond the scope of ggplot all together.

[chuxinyuan]

declain a function like this: pmatrix(list(p1,p2,p3,p4,p5) ,nrow=2) will product 2 rows，3+2，I think it is simple to understand，because we can reference matrix(),but it maybe need more args to control inch size.

[thomasp85]

@chuxinyuan that is basically the patchwork::wrap_plots() function