Clarification of requirement of bipartiteness for HeteroConv wrapper

[90degs2infty]

Hi everyone!

First and foremost thanks for the tremendous work on this library! 🙏

I'm reaching out to you because I need some clarification of the docs on HeteroConv "at the conceptual" level. Its documentation on the convs parameter reads:

convs (Dict[Tuple[str, str, str], MessagePassing]) – A dictionary holding a bipartite MessagePassing layer for each individual edge type.

Notice the requirement of the passed MessagePassing layers to be bipartite.

Suppose I have an heterogeneous graph with multiple node and edge types where one of the edge types connects nodes of a single type (i.e. edge_key[0] == edge_key[-1]). For example, extending the example from the docs on HeteroConv, suppose we introduced another edge type tracking citations at the author level: ("author", "cites", "author"). Such an edge type may well include self-loops (e.g. an author self-citing previous work) or in general odd-length cycles (which to my understanding of wikipedia renders the graph non-bipartite).

Does the HeteroConv wrapper support such graphs or is the requirement of bipartite MessagePassing layers intended to prevent users from applying HeteroConv to such graphs? More specifically, I wonder if explicitly encoding self-loops in the graph "by hand" is supported by HeteroConv - despite the graph not being bipartite in this case (again, to my understanding of wikipedia 😅 ).

I guess part of my confusion stems from discussion #4637 and the check in check_add_self_loops introduced in #4647 . I do get that self-loops are not well-defined for edges connecting different node types (linking a node of source type to itself while at the same time enforcing edges to point to a different destination type is simply not possible). In the discussion and in the implemented check, the term bipartite and the notion of "edges connecting nodes of different types" are used somewhat interchangeably.

Sorry, if this question appears overly pedantic. I'm just trying to avoid potential bugs in my usage of HeteroConv 😇 Thanks for any clarification!