Feature request: New argument `axis.titles` for facet plots

[jbengler]

Hi Teun,
I would like to suggest a new argument axis.titles to control the display of axis titles in facet plots. In a way this would complement the existing arguments axes and axes.labels.

Best
Broder

[teunbrand]

Hi Broder, would you suggest that all axis titles along a dimension have the same title? If not, how would one control the set the titles?

[jbengler]

The same title. I am aware that this will basically duplicate information that is already present. However, one would be able to give the facets the look and feel of individual plots.

[teunbrand]

Thanks for elaborating! I agree that this can be convenient to have, but at this point I'd just be inclined to use patchwork to get the layout you're after. Alternatively, you (or anyone who builds extensions) could make a facet extension where you override the Facet$draw_labels() method. The extension system was build so that ggplot2 doesn't have to implement every kind of option :)

[jbengler]

Thank you for thinking about it! I will follow your suggestion and have a look into Facet$draw_labels().

[jbengler]

I gave it a shot with Facet$draw_labels(). However, the gaps between panels can vary in width and therefore I cannot just make three equally sized cols to draw the labels.

Maybe I can get the exact positioning of the panels from somewhere?

Or I might even need to dive into the code that draws the panels itself. In this case I would probably go for FacetWrap$attach_axes() and include the labels there. However, this code is a little intimidating 😂

[teunbrand]

You can use panel_cols() and panel_rows() to find the correct cells in the gtable structure. However, it looks like these rectangles were drawn using freehand? If that is the case, then it might just be easier to work with the gtable structure rather than parallel to it.

However, this code is a little intimidating

It used to be part of a larger, even more intimidating chunk of code, so now it is just a smaller chunk of intimidating code.

[jbengler]

However, it looks like these rectangles were drawn using freehand?

Nope, they were done with code:

  xlab_height_bottom <- grobHeight(labels$x[[2]])
  tab <- gtable(unit(rep(1, 3), "null"), xlab_height_bottom)
  tab <- gtable_add_grob(tab, rectGrob(gp = gpar(fill = "#56B4E9")), t = 1, l = 1)
  tab <- gtable_add_grob(tab, rectGrob(gp = gpar(fill = "#F5C710")), t = 1, l = 2)
  tab <- gtable_add_grob(tab, rectGrob(gp = gpar(fill = "#D55E00")), t = 1, l = 3)
  panels <- gtable_add_rows(panels, xlab_height_bottom, pos = -1)
  panels <- gtable_add_grob(panels, tab, name = "xlab-b",
                            l = panel_dim$l, r = panel_dim$r, t = -1, clip = "off")

You can use panel_cols() and panel_rows() to find the correct cells in the gtable structure.

Thanks for this great tip!

Slowly getting closer...

It used to be part of a larger, even more intimidating chunk of code, so now it is just a smaller chunk of intimidating code.

Thankfully, I did not have to go in there!

[jbengler]

How would you check for empty panels that do not need axis titles?
Like to bottom right one in the plot above.

[teunbrand]

Ah it seems like you're getting to the juicy bits of why the faceting code is so intimidating 😅 The layout data frame should have ROW and COL variables that should allow you to deduce which panels are empty.

[jbengler]

Okay, with some try and error I arrived at this somewhat cryptic solution. It seems to work well for in cases where scales = "free", which is true for my use case.

Would you consider this a reasonable starting point to move further?

FacetWrap$draw_labels <- function(panels, layout, x_scales, y_scales, ranges, coord, data, theme, labels, params) {
  panel_dim <-  find_panel(panels)
  my_rows <- panel_rows(panels)
  my_cols <- panel_cols(panels)

  # x axis titles (bottom)
  xlab_height_bottom <- grobHeight(labels$x[[2]])
  title_rows <- my_rows - 4 + 0:(nrow(my_rows)-1)
  title_rows <- rbind(title_rows[-1, ], title_rows[1, ])
  for (i in 1:nrow(title_rows)) {
    panels <- gtable_add_rows(panels, xlab_height_bottom, pos = title_rows$t[i])
    for (k in 1:nrow(my_cols)) {
      if (any(layout$ROW == i & layout$COL == k)) {
        panels <- gtable_add_grob(panels, labels$x[[2]], name = "xlab-b",
                                  l = my_cols$l[k], r = my_cols$r[k], t = title_rows$t[i], clip = "off")
      }
    }
  }

  panel_dim <- find_panel(panels)
  my_rows <- panel_rows(panels)
  my_cols <- panel_cols(panels)

  # y axis labels (left)
  ylab_width_left <- grobWidth(labels$y[[1]])
  title_cols <- my_cols - 2 + 0:(nrow(my_cols)-1)
  for (i in 1: nrow(title_cols)) {
    panels <- gtable_add_cols(panels, ylab_width_left, pos = title_cols$l[i])
    for (k in 1: nrow(my_rows)) {
      if (any(layout$ROW == k & layout$COL == i)) {
        panels <- gtable_add_grob(panels, labels$y[[1]], name = "ylab-l",
                                  l = title_cols$l[i]+1, b = my_rows$b[k], t = my_rows$t[k], clip = "off")
      }
    }
  }
  panels
}

[jbengler]

One more thought:
I found it quite easy to build my own FacetSplit by using FacetWrap as template and only overwrite draw_labels()

However, the corresponding function facet_split() is basically the same as facet_wrap() only using FacetSplit instead of FacetWrap as the output ggproto, and I found no other way then repeating all the code inside facet_wrap() plus having to deal with some private ggplot2 functions (check_labeller(), compact_facets(), check_number_whole()) on the way.

Am I maybe missing a more elegant way to do this?

[teunbrand]

Would you consider this a reasonable starting point to move further?

The only issue I can spot on a glance is that you're modifying the FacetWrap$draw_labels() method directly, instead of overriding that method in a subclass. The ggproto classes do not have copy-on-modify semantics like most R objects so subclassing is necessary.

Am I maybe missing a more elegant way to do this?

I wouldn't say 'elegant', but I and some others sometimes just use the facet_wrap() constructor to steal the (parsed) parameters and forward it to our custom classes. An example: https://github.com/thomasp85/ggforce/blob/main/R/facet_row.R

[jbengler]

The ggproto classes do not have copy-on-modify semantics like most R objects so subclassing is necessary.

I now have my own FacetSplit that inherits from FacetWrap 👍

I wouldn't say 'elegant', but I and some others sometimes just use the facet_wrap() constructor to steal the (parsed) parameters and forward it to our custom classes.

I still have to wrap my head around this one, but it looks like a very powerful approach.

Thanks for all your great tips!