Better cooperation of as_data_frame() and graph_from_data_frame() ?

[mbojan]

It would be great to have the following functionality back:

dlist <- as_data_frame(g, what="both")
graph_from_data_frame(dlist$edges, vertices=dlist$vertices)

In other words, using (parts of) output of as_data_frame in graph_from_data_frame directly. I recall it was possible before recent API changes. Above will work if g has vertex names, but it won't if it does not. See the following reprex:

library(igraph)
#> 
#> Attaching package: 'igraph'
#> The following objects are masked from 'package:stats':
#> 
#>     decompose, spectrum
#> The following object is masked from 'package:base':
#> 
#>     union
g <- make_graph(~ A, B -- C, C -- D)
V(g)$a <- letters[1:4]
g1 <- delete_vertex_attr(g, "name")

# 1
dlist <- as_data_frame(g, what="both")
# Works if `g` has names
r <- graph_from_data_frame(dlist$edges, vertices=dlist$vertices, directed=is_directed(g))
identical_graphs(g, r) # Why?
#> [1] FALSE
identical(dlist, as_data_frame(r, what="both"))
#> [1] TRUE

# 2
dlist1 <- as_data_frame(g1, what="both")
# Won't work if `g` does not have names
graph_from_data_frame(dlist1$edges, vertices=dlist1$vertices, directed=is_directed(g1))
#> Error in graph_from_data_frame(dlist1$edges, vertices = dlist1$vertices, : Some vertex names in edge list are not listed in vertex data frame

In particular:

1. I think the main reason that this connection is lost is that as_data_frame( . , what="vertices") does not include the vertex IDs as the first column anymore.
2. I don't understand why g and r are not identical_graphs in # 1. Both edge and vertex data frames are identical. Am I missing something?

[clhunsen]

2. I don't understand why g and r are not identical_graphs in #1. Both edge and vertex data frames are identical. Am I missing something?

We have the same problem in our scripts. This seems weird as the same function call to identical_graphs works in other places.

What can we do to help debugging this?

---

@bockthom, @hechtlC: FYI.

[clhunsen]

Quick update to this: When I try a cascade of directed = TRUE and as.undirected(r), it works...

g <- make_graph(~ A, B -- C, C -- D)
V(g)$a <- letters[1:4]
dlist <- as_data_frame(g, what="both")
r <- graph_from_data_frame(dlist$edges, vertices=dlist$vertices, directed=TRUE)
r <- as.undirected(r)
identical_graphs(g, r)
# [1] TRUE

I hope this helps identifying the problem.

[clhunsen]

Any news on this?

[krlmlr]

Thanks, confirmed.

The first problem seems to be due to a bogus difference between an empty vector and NULL, need to take a look.

The second problem can be fixed with a more careful implementation of graph_from_data_frame() .

library(igraph, warn.conflicts = FALSE)

g <- make_graph(~A, B - -C, C - -D)
V(g)$a <- letters[1:4]
g1 <- delete_vertex_attr(g, "name")

# 1
dlist <- as_data_frame(g, what = "both")
# Works if `g` has names
r <- graph_from_data_frame(dlist$edges, vertices = dlist$vertices, directed = is_directed(g))
identical_graphs(g, r) # Why?
#> [1] FALSE
waldo::compare(unclass(g)[1:9], unclass(r)[1:9])
#> `names(old[[9]][[4]])` is a character vector ()
#> `names(new[[9]][[4]])` is absent

# 2
dlist1 <- as_data_frame(g1, what = "both")
# Won't work if `g` does not have names
graph_from_data_frame(dlist1$edges, vertices = dlist1$vertices, directed = is_directed(g1))
#> Error in graph_from_data_frame(dlist1$edges, vertices = dlist1$vertices, : Some vertex names in edge list are not listed in vertex data frame

^{Created on 2023-03-31 with reprex v2.0.2}

[krlmlr]

The first problem is now tracked in #756.

[krlmlr]

For the second problem, there are a lot of tests, and probably downstream packages, that rely on graph_from_data_frame() coercing integer vertices to strings. Effectively, this makes it difficult to have graph_from_data_frame() construct an unnamed graph, which seems to be a prerequisite for the roundtrip.

After having tinkered with it, I'd like to change the following in graph_from_data_frame():

• Only accept data frames and character vectors in the vertices argument
• Explicitly watch out for a $name attribute in vertices
• Interpret numeric inputs as vertex IDs in the first two columns of d

We can also extend as_data_frame() to offer an option to always return the edges as numeric vertex IDs.

If we want to be backward-compatible, we could request numeric vertex IDs in the edges to be wrapped by I() . It would be cleanest to make this a breaking change though, but we need to assess the impact first.