STAT 436 (Spring 2023): Graphical Queries

Some of the most sophisticated interactive data visualizations are based on the idea that user queries can themselves be defined visually. For example, to select a date range, we could directly interact with a time series plot, rather than relying on a slider input. Or, instead of a long dropdown menu of items, a user could select items by clicking on bars in a bar plot. There are many variations of this idea, but they all leverage graphical (rather than textual) displays to define queries. The advantage of this approach is that it increases information density – the selection inputs themselves encode data.
To implement this in Shiny, we first need a way of registering user interactions on plots themselves. We will consider two types of plot interaction mechanisms: clicks and brushes. These can be specified by adding click or brush events to plotOutput objects.
This creates a UI with a single plot on which we will be able to track user clicks,
```
ui <- fluidPage(
   plotOutput("plot", click = "plot_click")
 )
```
Here, plot_click is an ID that can be used as input$plot_click in the server. We could name it however we want, but we need to be consistent across the UI and server (just like ordinary, non-graphical inputs).
Before, we just needed to place the input$id items within render and reactive server components, and the associated outputs would automatically know to redraw each time the value of any input was changed. Clicks are treated slightly differently. We have to both (a) recognize when a click event has occurred and (b) extract relevant information about what the click was referring to.
For (a), we generally use observeEvent,
```
observeEvent(
  input$plot_click,
  ... things to do when the plot is clicked ...
)
```
This piece of code will be run anytime the plot is clicked.
For (b), we can use the nearPoints helper function. Suppose the plot was made using the data.frame x. Then
```
nearPoints(x, input$click)
```
will return the samples in x that are close to the clicked location. We will often use a variant of this code that doesn’t just return the closeby samples – it returns all samples, along with their distance from the clicked location,
```
nearPoints(x, input$click, allRows = TRUE, addDist = TRUE)
```
We are almost ready to build a visualization whose outputs respond to graphical queries. Suppose we want a scatterplot where point sizes update according to their distance from the user’s click. Everytime the plot is clicked, we need to update the set of distances between samples and the clicked point. We then need to rerender the plot to reflect the new distances. This logic is captured by the block below,
```
server <- function(input, output) {
   dist <- reactiveVal(rep(1, nrow(x)))
   observeEvent(
     input$plot_click,
     dist(reset_dist(x, input$plot_click))
   )
   
   output$plot <- renderPlot({
     scatter(x, dist())
   })
 }
```
The code above uses one new concept, the reactiveVal on the first line of the function. It is a variable that doesn’t directly depend on any inputs, which can become a source node for downstream reactive and render nodes in the reactive graph. Anytime the variable’s value is changed, all downstream nodes will be recomputed. A very common pattern is use an observeEvent to update a reactiveVal every time a graphical query is performed. Any plots that depend on this value will then be updated. For example,
```
val <- reactiveVal(initial_val) # initialize the reactive value

observeEvent(
  ...some input event...
  ...do some computation...
  val(new_value) # update val to new_val
)

# runs each time the reactiveVal changes
renderPlot({
  val() # get the current value of the reactive value
})
```

So, revisiting the dist in the earlier code block, we see that it is initialized as a vector of 1’s whose length is equal to the number of rows of x. Everytime the plot is clicked, we update the value of dist according to the function reset_dist. Finally, the changed value of dist triggers a rerun of renderPlot. Let’s look at the full application in action. It makes a scatterplot using the cars dataset and resizes points every time the plot is clicked.

library(tidyverse)
 library(shiny)
 
 # wrapper to get the distances from points to clicks
 reset_dist <- function(x, click) {
   nearPoints(x, click, allRows = TRUE, addDist = TRUE)$dist_
 }
 
 # scatterplot plot with point size dependent on click location
 scatter <- function(x, dists) {
   x %>%
     mutate(dist = dists) %>%
     ggplot() +
     geom_point(aes(mpg, hp, size = dist)) +
     scale_size(range = c(6, 1))
 }
 
 ui <- fluidPage(
   plotOutput("plot", click = "plot_click")
 )
 
 server <- function(input, output) {
   dist <- reactiveVal(rep(1, nrow(mtcars)))
   observeEvent(
     input$plot_click,
     dist(reset_dist(mtcars, input$plot_click))
   )
   
   output$plot <- renderPlot(scatter(mtcars, dist()))
 }
 
 shinyApp(ui, server)

The reset_dist function uses nearPoints to compute the distance between each sample and the plot, each time the plot is clicked. The associated reactive value dist gets changed, which triggers scatterplot to run, and it is encoded using size in the downstream ggplot2 figure.

We can make the plot more interesting by outputting a table showing the original dataset. Using the same dist() call, we can sort the table by distance each time the plot is clicked.

library(tidyverse)
 library(shiny)
 mtcars <- add_rownames(mtcars)
 
 reset_dist <- function(x, click) {
   nearPoints(x, click, allRows = TRUE, addDist = TRUE)$dist_
 }
 
 scatter <- function(x, dists) {
   x %>%
     mutate(dist = dists) %>%
     ggplot() +
     geom_point(aes(mpg, hp, size = dist)) +
     scale_size(range = c(6, 1))
 }
 
 ui <- fluidPage(
   plotOutput("plot", click = "plot_click"),
   dataTableOutput("table")
 )
 
 server <- function(input, output) {
   dist <- reactiveVal(rep(1, nrow(mtcars)))
   observeEvent(
     input$plot_click,
     dist(reset_dist(mtcars, input$plot_click))
   )
   
   output$plot <- renderPlot(scatter(mtcars, dist()))
   output$table <- renderDataTable({
     mtcars %>%
       mutate(dist = dist()) %>%
       arrange(dist)
   })
 }
 
 shinyApp(ui, server)

Graphical Queries - Click Events