If the natural `ggplot2`

equivalent to nodes is
`geom_point()`

, then surely the equivalent to edges must be
`geom_segment()`

? Well, sort of, but there’s a bit more to it
than that.

While nodes are the sensible, mature, and predictably geoms, edges are the edgy (sorry), younger cousins that pushes the boundaries. To put it bluntly:

On the ggraph savannah you definitely want to be an edge!

`geom_edge_*()`

familyWhile the introduction might feel a bit over-the-top it is entirely
true. An edge is an abstract concept denoting a relationship between two
entities. A straight line is simply just one of many ways this
relationship can be visualised. As we saw when discussing nodes sometimes it is not drawn at all
but impied using containment or position (treemap, circle packing, and
partition layouts), but more often it is shown using a line of some
sort. This use-case is handled by the large family of edge geoms
provided in `ggraph`

. Some of the edges are general while
others are dedicated to specific layouts. Let’s creates some graphs for
illustrative purposes first:

```
library(ggraph)
library(tidygraph)
library(purrr)
library(rlang)
set_graph_style(plot_margin = margin(1,1,1,1))
hierarchy <- as_tbl_graph(hclust(dist(iris[, 1:4]))) |>
mutate(Class = map_bfs_back_chr(node_is_root(), .f = function(node, path, ...) {
if (leaf[node]) {
as.character(iris$Species[as.integer(label[node])])
} else {
species <- unique(unlist(path$result))
if (length(species) == 1) {
species
} else {
NA_character_
}
}
}))
hairball <- as_tbl_graph(highschool) |>
mutate(
year_pop = map_local(mode = 'in', .f = function(neighborhood, ...) {
neighborhood %E>% pull(year) |> table() |> sort(decreasing = TRUE)
}),
pop_devel = map_chr(year_pop, function(pop) {
if (length(pop) == 0 || length(unique(pop)) == 1) return('unchanged')
switch(names(pop)[which.max(pop)],
'1957' = 'decreased',
'1958' = 'increased')
}),
popularity = map_dbl(year_pop, ~ .[1]) %|% 0
) |>
activate(edges) |>
mutate(year = as.character(year))
```

While you don’t have to use a straight line for edges it is certainly
possible and `geom_edge_link()`

is here to serve your
needs:

There’s really not much more to it — every edge is simply a straight line between the terminal nodes. Moving on…

Sometimes the graph is not simple, i.e. it has multiple edges between
the same nodes. Using links is a bad choice here because edges will
overlap and the viewer will be unable to discover parallel edges.
`geom_edge_fan()`

got you covered here. If there are no
parallel edges it behaves like `geom_edge_link()`

and draws a
straight line, but if parallel edges exists it will spread them out as
arcs with different curvature. Parallel edges will be sorted by
directionality prior to plotting so edges flowing in the same direction
will be plotted together:

An alternative to `geom_edge_fan()`

is
`geom_edge_parallel()`

. It will draw edges as straight lines
but in the case of multi-edges it will offset each edge a bit so they
run parallel to each other. As with `geom_edge_fan()`

the
edges will be sorted by direction first. The offset is done at draw time
and will thus remain constant even during resizing:

Loops cannot be shown with regular edges as they have no length. A
dedicated `geom_edge_loop()`

exists for these cases:

```
# let's make some of the student love themselves
loopy_hairball <- hairball |>
bind_edges(tibble::tibble(from = 1:5, to = 1:5, year = rep('1957', 5)))
ggraph(loopy_hairball, layout = 'stress') +
geom_edge_link(aes(colour = year), alpha = 0.25) +
geom_edge_loop(aes(colour = year))
```

The direction, span, and strength of the loop can all be controlled, but in general loops will add a lot of visual clutter to your plot unless the graph is very simple.

This one is definitely strange, and I’m unsure of it’s usefulness,
but it is here and it deserves an introduction. Consider the case where
it is of interest to see which types of edges dominates certain areas of
the graph. You can colour the edges, but edges can tend to get
overplotted, thus reducing readability. `geom_edge_density()`

lets you add a shading to your plot based on the density of edges in a
certain area:

```
ggraph(hairball, layout = 'stress') +
geom_edge_density(aes(fill = year)) +
geom_edge_link(alpha = 0.25)
```

```
## Warning: The following aesthetics were dropped during statistical transformation: xend
## and yend.
## 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?
```

While some insists that curved edges should be used in standard
*“hairball”* graph visualisations it really is a poor choice, as
it increases overplotting and decreases interpretability for virtually
no gain (unless complexity is your thing). That doesn’t mean arcs have
no use in graph visualizations. Linear and circular layouts can benefit
greatly from them and `geom_edge_arc()`

is provided precisely
for this scenario:

Arcs behave differently in circular layouts as they will always bend
towards the center no matter the direction of the edge (the same thing
can be achieved in a linear layout by setting
`fold = TRUE`

).

Edge bundling is a technique to reduce clutter in a network visualization by bundling edges that flows in the same direction. There are various ways of doing this, many with heavy computational cost and the potential to mislead. The technique were initially confined to connections between nodes with a hierarchical structure but has been expanded to general graphs. ggraph provides 3 different bundling geoms with various up- and downsides.

This is perhaps the most classic. It treats the edges as an array of points with the propensity to attract each other if edges are parallel. It suffers from bad performance (though the edge bundling geoms uses memoisation to avoid recomputations) and can also be misleading as it doesn’t use the underlying topology of the graph to determine if edges should be bundled, only whether they are parallel.

`## Using "stress" as default layout`

An alternative is to let the edges follow the shortest paths rather than attract each other. This means the topology is being used in the bundling and in theory lead to less misleading results. It also has the upside of being faster. The algorithm is iterative so that if an edge has been bundled it is deleted from the graph where the shortest path is being searched in. In this way the edges naturally converge towards a few “highways”.

`## Using "stress" as default layout`