Signed topological graph learning

Investigate signed topological graph learning by developing modern pre-training, prompting, or graph neural network representation-learning frameworks beyond existing homological characterizations of structural balance.

Background

The paper explains that most topological graph-learning methods assume unsigned or merely weighted graphs. For signed graphs, prior work is described as mainly characterizing structural balance through simplicial homology and cohomology, without providing contemporary pre-training, prompting, or graph neural network representation-learning methods. Consequently, the broader problem of learning useful representations from signed topology remains unresolved. The proposed TopoSIGN framework addresses signed topology through a signed degree-vector Dowker filtration, but the paper identifies signed topological graph learning as a broader research area that is still open.