Test suite fails when using "igraph" versions >= 1.4.0

[bockthom]

In igraph version 1.4.0, they have introduced a breaking change regarding attribute types (see here: https://github.com/igraph/rigraph/blob/main/NEWS.md#igraph-140). This breaking change in igraph 1.4.0 breaks our test suite seriously, hundreds of tests fail.

igraph 1.4.0 was released on 22nd of February 2023. There are multiple reasons how the recently introduced breaking changes of igraph 1.4.0 affect our test suite. For the moment, I just looked at one failing test and spotted two of these reasons, but there can also be more than these two, of course:

---

(1) Using the AsIs class to generate the expected test output lets tests fail

igraph now respects all types and classes (which have been ignored before). That is, creating data.frames containing I(...) calls and creating networks from the data.frame via graph.data.frame now creates networks that are different than the networks coronet creates, due the AsIs class introduced by I(...) which has been ignored in previous igraph versions.
However, this can be easily fixed by calling unclass for each data.frame column that has been created via I(...).

Example: Consider the following test for author networks which uses I(...) to create the expected data.frame.

coronet/tests/test-networks-author.R

Lines 425 to 443 in b7db5cd

	data = data.frame(
	from = c("Björn", "Olaf", "Olaf", "Karl"),
	to = c("Olaf", "Karl", "Thomas", "Thomas"),
	date = I(list(c(1468339139, 1468339245), c(1468339541, 1468339570), c(1468339541, 1468339592),
	c(1468339570, 1468339592))),
	artifact.type = I(list(c("Feature", "Feature"), c("Feature", "Feature"), c("Feature", "Feature"),
	c("Feature", "Feature"))),
	hash = I(list(
	c("72c8dd25d3dd6d18f46e2b26a5f5b1e2e8dc28d0", "5a5ec9675e98187e1e92561e1888aa6f04faa338"),
	c("3a0ed78458b3976243db6829f63eba3eead26774", "1143db502761379c2bfcecc2007fc34282e7ee61"),
	c("3a0ed78458b3976243db6829f63eba3eead26774", "0a1a5c523d835459c42f33e863623138555e2526"),
	c("1143db502761379c2bfcecc2007fc34282e7ee61", "0a1a5c523d835459c42f33e863623138555e2526"))),
	file = I(list(c("test.c", "test.c"), c("test2.c", "test3.c"), c("test2.c", "test2.c"), c("test3.c", "test2.c"))),
	artifact = I(list(c("A", "A"), c("Base_Feature", "Base_Feature"), c("Base_Feature", "Base_Feature"),
	c("Base_Feature", "Base_Feature"))),
	weight = 2,
	type = TYPE.EDGES.INTRA,
	relation = "cochange"
	)

For each of the columns that uses I(...) we just need to call unclass afterwards. So, in the example above, we need to add the following lines below the data.frame creation:

 data[["date"]] = unclass(data[["date"]])
 data[["artifact.type"]] = unclass(data[["artifact.type"]])
 data[["hash"]] = unclass(data[["hash"]])
 data[["file"]] = unclass(data[["file"]])
 data[["artifact"]] = unclass(data[["artifact"]])

This fix does not affect previous versions of igraph. Hence, for all igraph versions, this problem can be fixed by just adding these unclass calls where applicable.

---

(2) igraph now respects types in edge attributes

In previous versions of igraph, types of newly set edge attributes where ignored. For example, the date attribute was introduced as an attribute of type "numeric". Even when the actual attribute was set as an "POSIXct", the date attribute was still stored as being of type "numeric". However, now the types are respected and dates are stored as "POSIXct". In most of our tests, this is not a problem, as the created data.frame for the expected network already correctly converts the dates to "POSIXct" - so our expected data is correct and the tests formerly worked though because this type was just ignored. However, there are tests which have a weird behavior:

Let's have a look at the author-network creation test already discussed above:
Example: Consider the following test for author networks which uses I(...) to create the expected data.frame.

coronet/tests/test-networks-author.R

Lines 428 to 429 in b7db5cd

	date = I(list(c(1468339139, 1468339245), c(1468339541, 1468339570), c(1468339541, 1468339592),
	c(1468339570, 1468339592))),

For whatever reason, in this test, unix timestamps are used instead of dates in string format. Therefore, this test fails with igraph >= 1.4.0, as unix timestamps are of type "numeric" and not of type "POSIXct". This is easily fixable - either by calling get.date.from.unix.timestamp or by using the string representation of the date, as in other tests. However, with this fix, the test does not work any more for igraph versions < 1.4.0. So, we need a solution here that works for both recent and older igraph versions.
@hechtlC As you have authored this test, do you remember why you have used the unix timestamps here?
We need to find out why this test behaves different than all the other tests - either igraph did (and still does) some inconsistent type processing, or we somehow do some inconsistent processing in coronet.

---

This issue affects all versions of coronet from version 3.1 on up until now (i.e., version 4.2).

[bockthom]

Short summary: @hechtlC and I had decided to fix the test in coronet/tests/test-networks-author.R (see comment above) in such a way that we extract the class of the date attribute from the built network and choose the right date conversion function for this single test afterwards accordingly. This way, we make sure that this test passes independent of the used igraph version (i.e., igraph < 1.4.0 or igraph >= 1.4.0).

---

In addition to the above described problems, a couple of other problems arose in relation to the new attribute class handling as of igraph version 1.4.0. I have already implemented some fixes for them.
For all the problems that arose in relation to this breaking change of igraph, I will open a pull request soon that fixes all the related problems. See the corresponding pull request for more details.