Available interactions

There are many ways of interacting with plotscaper figures, both client-side with the figure directly, and from within a running R session.

This vignette isn’t meant to provide an exhaustive overview - I’m terrible at book-keeping, so I may forget to add/remove descriptions of some features as time goes on (but I promise to try my best not to). Meanwhile, you may be better off approaching this vignette as a compendium of the most important/stable features. If you have any questions, please consult the package documentation or hit me up via email.

Linked selection: Bread and butter

This is one of the main features of plotscaper (or indeed, most systems for interactive data visualization). All of the plots in a plotscaper figure are linked, so clicking or clicking-and-dragging can be used to select objects in one plot, and the corresponding cases are then highlighted across all the other plots.

By default, selection is transient, meaning that if you make a selection and then click anywhere else in the figure, the selection disappears. You can make the selection permanent and assign cases to permanent groups by holding the key 1, 2, or 3 while doing the selection. Permanent selection is only removed by double-clicking.

This is all client-side interactions (clicking buttons and pressing keys). You can achieve the same effect from R by using the select_cases and assign_cases functions:

library(plotscaper)

schema <- create_schema(mtcars) |> 
  add_scatterplot(c("wt", "mpg")) |> 
  add_barplot(c("cyl"))

schema |> select_cases(1:10) |> render()

schema |> 
  assign_cases(1:10) |> # Use e.g. 'group = 2' to assign to a different group
  render()

Additionally, if you’re inside a running R session, you can also use the selected_cases and assigned_cases to query the current selection status. The document you’re reading is a static HTML, therefore these functions will not work here.

Transient and permanent selection combine into a Cartesian product (just a fancy way of saying “all possible combinations”), so you can have up to 8 distinct selection groups present in a figure at a time: base, base + transient, group 1, group 1 + transient, etc…

Now, it might not be the best idea to have 8 different selection groups present in a plot at a single time - from static data visualization, we know that the number of distinct groups we can efficiently perceive based on color is limited, so once we go past, say, three we really start to strain our visual system.

However, combining transient selection with permanent selection can be useful for quickly answering queries about nested subsets of our data. For example, here’s how we can figure out how many of the 8-cylinder cars are also automatic:

schema |> 
  assign_cases(which(mtcars$cyl == 8)) |>
  select_cases(which(mtcars$am == 1)) |>
  render()

There are some additional considerations about linked selection, see the algebra vignette.

Zooming, panning, and scales

Zooming

Once we’ve made a selection, we can zoom into the selected region by pressing the zoom key (Z by default). We can zoom into a plot multiple times. To revert the zoom, we can either do a full reset of the figure (R key; will get rid off other temporary state as well, including histogram binwidth/anchor changes, barplot reordering, etc…), or “pop” one level of zoom (X key).

Server-side, we can use the zoom function to zoom into a specific region of a plot. For example, here’s how we can zoom into the middle 1/4 of the scatterplot:

schema |> zoom("plot1", c(0.25, 0.25, 0.75, 0.75)) |> render()

We can think of the coordinates as a rectangle starting at the 25th percentiles and ending at the 75th percentiles, of both the x- and the y-axis. On each axis, this covers exactly 50% of plot area, and since we are doing this along both axes, the resulting size of the region is \(0.5 ^2 = 0.25\) or 25% (or 1/4).

By default, the zoom function uses the pct units, meaning that the coordinates we provide (vector of rectangle corners: x0, y0, x1, y1) are interpreted as percentages. We can change the units argument to abs (absolute pixels) or data (data coordinates; continuous scales only):

schema |> zoom("plot1", c(0, 0, 10, 50), units = "data") |> render()

Notice also that we need to provide the functions above a plot identifier as the second argument (first if we are using a pipe |>). The plot identifier is always in the form of "[prefix][numeric suffix]", where the prefix can either be generic "plot" or it can refer to a specific plot type, such as "scatter", "bar", or "histo". The suffix identifies the plot by a position in the figure, in the left-to-right, top-to-bottom order. For example, the "plot1" identifier identifies the left-top-most plot in the figure, "scatter1" identifies the left-top-most scatterplot, etc…

You can always retrieve a list of the available plot identifiers from a scene or schema by using the get_plot_ids function:

schema |> get_plot_ids()
#> [1] "scatter1" "bar1"

(the function returns the list of the identifiers indexed by the plot type)

Identifiers can also be shortened. You can read more about identifiers at ?id.

Panning

Panning is also supported out of the box. To pan a plot, simply click and hold the right mouse button and move the mouse.

Server-side, I haven’t provided a direct utility for panning. However, as I will show in the next section, we can achieve the same behavior either by using zoom or by manipulating scales.

Scales

To achieve the same result as panning panning manually, we can set the zero and one properties on the x- and y-scales. In simple terms, zero and one determine where the minimum and maximum of the data get mapped to on the codomain (screen position, in this case). For example, if zero = 0 and one = 1 and we’re mapping to a plotting region 800 pixels wide, then the minimum data value will get mapped to 0px and the maximum will get mapped to 800px. If zero = 0.1 and one = 0.9 (the default), the minimum data values gets mapped to 0.1 * 800 = 80 pixels and the maximum will get mapped to (800 * 0.9) = 720 pixels. Simple enough, right?

We can emulate panning by incrementing zero and one the same amount for a given scale. For example, here’s how we can “pan” both plots 50% of the viewport’s width to the right:

schema |> 
  set_scale("plot1", "x", zero = 0.6, one = 1.4) |> # zero/one = default value + 0.5
  set_scale("plot2", "x", zero = 0.6, one = 1.4) |> # ditto
  render()

The reason why zero and one have the value 0.1 and 0.9 by default is to prevent objects from overlapping the plotting region borders by expanding the scale. Here’s how it looks like when they’re actually set to the values 0 and 1:

schema |> 
  set_scale("plot1", "x", zero = 0, one = 1) |> 
  set_scale("plot1", "y", zero = 0, one = 1) |>
  # Barplot scales are a bit more complicated:
  # - for 'y' scale, zero is frozen (i.e. panning stretches bars up and down) 
  #   so we need to unfreeze this parameter to modify it
  # - for 'width', there is one more parameter that determines the width of the bar
  set_scale("plot2", "x", zero = 0, one = 1) |> 
  set_scale("plot2", "y", zero = 0, one = 1, unfreeze = TRUE) |>
  set_scale("plot2", "width", zero = 0, one = 1, mult = 1) |> 
  render()

These aren’t the best looking plots, as I’m sure you’d agree. Sensible default margins are very useful, most of the time.

Any changes to the scale parameters are temporary by default - if you press the reset key (R), they revert back to the defaults. If you want the changes to persist, add default = TRUE argument to the set_scale call.

Finally, if you want to inspect the values that a scale uses to translate data on the JavaScript side, you can use the get_scale function:

# NOT RUN - this only works in an active R session
scene |> get_scale("plot1", "x")

The output is a somewhat complicated list(). If you want to learn about how it works have a look at the documentation for ?get_scale.

We could also implement panning by specifying the minimum and maximum of the data scale explicitly. I generally don’t recommend doing this, since this only works for continuous scales and also goes against how zooming and panning are implemented in plotscaper: user interactions manipulate zero and one only and do not touch the data values on scales (even zoom(..., units = "data") does this). If you set the data limits explicitly, you may encounter some weird behavior (maybe not, but consider this a fair warning).

Anyway, I’d recommend using zoom(..., units = "data") but if you really want to mess with the data limits, you can do it:

schema |> 
  set_scale("plot1", "x", zero = 0, one = 1, min = 0, max = 10) |> 
  set_scale("plot1", "y", zero = 0, one = 1, min = 0, max = 50) |>
  render()

(notice also that the points stay constant size, whereas zoom shrinks/grows objects in response to zooming/unzooming)

Finally, we can also flip the direction of a scale like so:

schema |> set_scale("plot1", "y", direction = -1) |> render()

Discrete scales

We can also interactively reorder discrete scales. In a barplot, we can automatically sort bars by their value (O key).

Server-side, we can do this by setting the breaks argument (and we have more fine-grained control - we can order the bars any way we want):

schema |>
  set_scale("plot2", "x", breaks = c("6", "4", "8")) |>
  render()

Note: make sure that breaks match your factor levels exactly - failing to do so will result in a client-side error, which you will not see unless you open up the developer console (right click viewer panel + “inspect element”).

Size and opacity

We can manipulate the size and opacity of objects in the figure by using the grow/shrink keys (+/-) and the fade/unfade keys ([/]), respectively. Manipulating the size of objects actually works a bit differently depending on the type of the object and plot: in scatterplots, it modifies point area, in barplots, it modifies the bar widths, in histograms, it changes the binwidth (see more on that further below).

These interactions make the most sense to me on client-side, I may add some utilities to do them from the server-side too in the future.

Querying

You can query any geometric object in plotscaper by pressing and holding the query key (Q) and hovering over it.

Like manipulating size and opacity, querying does not have a server-side analog. This is because all of the data summaries exist client-side only. This is also what makes plotscaper relatively fast - everything is computed client-side, and communication with the R session happens only when necessary.

I could theoretically implement a way to query objects from the server-side, but coming up with a sensible API for this seems like a fairly complicated (and pointless) exercise - while hovering over a rectangle is a simple and unambiguous way to specify that you want to query that specific rectangle at those specific coordinates, how would you do the same in code? Maybe I’ll change my mind about this in the future.

For now, if you want the object summaries in R, you can always just subset the relevant rows (using selected_cases or assigned_cases to get the cases corresponding to the specific object) and compute the summaries yourself.

Parametric interaction: normalization, histograms

plotscaper also supports certain custom parametric interactions. For example, many of the aggregation plots such as barplot, histogram (1d or 2d), or fluctuation can be “normalized”, meaning that, for example, we can turn a barplot into a spineplot, which sets the total height of each bar to one and maps the count/statistic to the bars’ width. This is useful for accurately comparing proportions across bars of different heights. Client-side, we can do this by pressing the N key.

Server-side, we can call the normalize function. Compare the following two figures:

schema |> select_cases(5:15) |> render()

schema |> select_cases(5:15) |> normalize("plot2") |> render()

In the second figure, it’s much easier to see that the proportion of the selected cases for 6 and 8 cylinders is the same.

Histogram binwidths and anchors

Currently, the other parameters than can be interactively manipulated are histogram binwidth and anchors. This applies to both 1D and 2D histograms.

On the client-side you can use the grow/shrink keys (+/-) to change the size of the bins, and the increment/decrement anchor keys ('/;) to change the anchor. For 2D histograms, you can only change the binwidth, and this happens for both axes simultaneously; it didn’t seem to make sense to change anchor on both axes simultaneously, and I didn’t want to complicate the default interface by adding more keys. If you want fine-grained control over the 2D histogram, use the server-side functions.

On that note, server-side, you can use the set-parameters function:

create_schema(mtcars) |> 
  add_scatterplot(c("wt", "mpg")) |>
  add_histogram(c("wt")) |> 
  set_parameters("plot1", width = 1, anchor = 0.5) |>
  render()

(notice that the boundaries of the bins are aligned with the midpoints of the axis labels)

For 2D histograms, we need to specify which axis we want to set the parameters to:

names(airquality) <- c("ozone", "solar radiation", "wind", 
                       "temperature", "month", "day")

create_schema(airquality) |>
  add_histogram2d(c("solar radiation", "ozone")) |>
  add_pcoords(names(airquality)) |>
  set_parameters("plot1", width_x = 30, width_y = 15, 
                  anchor_x = 0, anchor_y = 0) |>
  render()
#> Warning in create_schema(airquality): Removed 42 rows with missing values from
#> the data