Support color mapping for continuous variable in scatter plots

[Ductmonkey]

GGpairs supports mapping color to categorical variables but it seems not for continuous variables. This could make the visualization much more powerful to help expose 3-dimensional correlations (or lack thereof).

Is it possible to allow this for panels where it would be sensible (i.e. scatter plot) and leave other panels black?

library(GGally)

# works fine for categorical variables
ggpairs(iris, columns = 1:4, aes(color = Species))

# fails on continuous variables
ggpairs(iris, columns = 1:4, aes(color = Sepal.Length))
#> Warning: The following aesthetics were dropped during statistical transformation: fill
#> ℹ This can happen when ggplot fails to infer the correct grouping structure in
#>   the data.
#> ℹ Did you forget to specify a `group` aesthetic or to convert a numerical
#>   variable into a factor?
#> Error in ggally_statistic(data = data, mapping = mapping, na.rm = na.rm, : `mapping` color column must be categorical, not numeric

^{Created on 2024-08-06 with reprex v2.0.2}

[rodmellot]

Hello. I am new to the package and together with @melanielacour and
under the supervision of [@katossky / @JMeyer31] we are willing to solve this
issue, as part of an assignment at the ENSAI school of statistical engineering
(France). We will try to be as autonomous as possible but we may ask for guidance from now and then. We have a solid background in statistics and a reasonable one in computer science. If any maintainer thinks this issue is too ambitious for a beginner, please tell us so and - if you have time - guide us to a project / issue were our help may be more useful.

[katossky]

@Ductmonkey can then @rodmellot and @melanielacour start with dealing with scatterplots only ? I guess that using discretisation of the colour scale would make filling the rest of the matrix just easier. As you mentioned, they would then leave the rest of the plot matrix as it is without any colouring variable.

The discretisation could be use, later, for diagonals : density ridges or colour density ranges. (However I am not sure the aes(colour = Sepal.Length) would be an intuitive way to call for such plot.)

Upper diagonals could have a line plot of smoothed out correlations along the conditional variable. (Here also, I doubt aes(colour = Sepal.Length) is be the best way to name call for for it.)

[melanielacour]

Hello,
After reviewing the package code and attempting to modify the ggpairs() function, we realized that using a call like (mapping = aes(Sepal.Length)) was problematic.

To obtain the desired plot, a call like the one below could help.

library(GGally)
library(ggplot2)

ggpairs(
  iris,                               # Data
  columns = 1:4,                      # Columns to plot (1:4 means all numeric variables)
  mapping = aes(colour = Species),    # Color by Species column
  title = "Pairwise Plots for Iris Dataset", # Title of the graph
  upper = list(continuous = "points", combo = "dot_no_facet"),  # Upper part: Points for continuous variables, dots for combos
  lower = list(continuous = "smooth", combo = "facetdensity"), # Lower part: Smooth lines for continuous, density for combos
  diag = list(continuous = "densityDiag", discrete = "barDiag") # Diagonal: Density for continuous, bar for discrete
)

However, we will propose a modification to obtain linear regression coefficients on the upper part and a discretized density on the diagonal.

[rodmellot]

Hello,
Here are the results we have with @melanielacour,
After trying to simplify the calls despite everything, due to mapping, it is necessary to keep the details on the different plots to be displayed. The code we propose allows decomposing the density according to the variable used in mapping and produces a result like the image that follows. Does this help you solve your problem?

NB : The lower part is optional. However, to avoid errors in the production of correlation values, which are the default choice, it is necessary to keep the upper part.

ggpairs(iris,
              columns = 1:4,
              mapping = aes(colour = Sepal.Length), 
              upper = list(continuous = "points"),
              lower = list(continuous = "smooth"),  
              diag = list(continuous = "ridgeDiag") )

Here is the link to our repository of our public fork: https://github.com/rodmellot/ggally_2.git.

[katossky]

Too me at least it looks like you solved the issue that you said you would.

There remains the issue with colouring the diagonal, but hopefully someone else will deal with this ?

[katossky]

Maybe the variable by which the dataset is discretised on the diagonal is not that clear though.

[Ductmonkey]

Thanks @rodmellot! I'm sorry I am just seeing this now. I realize the issue was probably that I can't map a continuous variable to the fill color of the density plots in the diagonal. If I specify appropriate geoms, this works fine with the original function in {GGally}:

library(ggplot2)
library(GGally)

ggpairs(iris, 
        columns = 1:4, 
        mapping = aes(colour = Sepal.Length), 
        upper = list(continuous = "points"),
        lower = list(continuous = "points"))