Allow for user-specified & cross-experiment-consistent column ordering

[katossky]

First thanks so much for the package. This is not criticism even though I am not really good at making it sound nice.

It is at the moment quite tedious to preserve the ordering of corr-generated plots in an other plot.

Say I want to try 2 different correlation measures, or two ways to pre-process my data, I could not found any straightforward way to plot the correlations in exactly the same variable order. I've just spent 1h30 trying to hack the resulting cor_df and ggplot object to get what I wanted without success. I had to build the plot from scratch, inspecting the autoplot code in order to reach my goal.

I see two ways forward :

1. add examples of how to do that simply (hopefully there is an undocumented easy way)
2. add an option to rearrange for user-specified order

---

Current strategy :

g <- iris |>
  select(-Species) |>
  corrr::correlate(quiet = TRUE) |>
  corrr::rearrange() |>
  corrr::shave() |>
  corrr::stretch() |>
  dplyr::mutate(
    x = factor(x, levels = unique(x)),
    y = factor(y, levels = unique(y))
  ) |>
  dplyr::filter(!is.na(r)) |>
  ggplot() + aes(x = x, y = y, fill = r) +
  geom_tile(color = "white") +
  scale_fill_viridis_c(limits = c(-1, 1)) +
  scale_x_discrete() +
  labs(x = NULL, y = NULL, fill = NULL) +
  coord_fixed() +
  theme_minimal() +
  theme(
    panel.grid = element_blank(), 
    axis.text.x = element_text(angle = 315, vjust = 1, hjust = 0)
  )
g

Now change the default correlation to spearman.

ordering <- levels(g$data$x)

iris |>
  select(-Species) |>
  corrr::correlate(quiet = TRUE, method = "spearman") |> # new !!!
  corrr::stretch() |>
  dplyr::mutate(
      x = factor(x, levels = ordering), # new !!!
      y = factor(y, levels = ordering)  # new, needed for filtering !!!
  ) |>
  dplyr::filter(!is.na(r), as.integer(x) < as.integer(y)) |> # new !!!
  dplyr::mutate(
      y = factor(y, levels = rev(ordering)) # back to standard plottting !!!
  ) |>
  ggplot() + aes(x = x, y = y, fill = r) +
  geom_tile(color = "white") +
  scale_fill_viridis_c(limits = c(-1, 1)) +
  scale_x_discrete() +
  labs(x = NULL, y = NULL, fill = NULL) +
  coord_fixed() +
  theme_minimal() +
  theme(
      panel.grid = element_blank(), 
      axis.text.x = element_text(angle = 315, vjust = 1, hjust = 0)
  )

Ideal strategy :

iris_corr <- iris |>
  select(-Species) |>
  corrr::correlate(quiet = TRUE) |>
  corrr::rearrange(method = "PCA")
cols <- setdiff(colnames(iris_corr), "term") # long term ordering for side-to-side comparisons
autoplot(iris_corr)

# then variations
iris |>
  select(-Species) |>
  corrr::correlate(quiet = TRUE, method = "pearsons") |>
  autoplot(ordering = cols)

iris |>
  mutate(Sepal.Length = ifelse(Sepal.Length > 0, 0, Sepal.Length)) |> # or whatever
  select(-Species) |>
  corrr::correlate(quiet = TRUE) |>
  autoplot(ordering = cols) # if this path is chosen, maybe should ordering supersede method but there might be edge cases in programming where this might create problems ?