·research·paper

Multi-graph former — a transformer over heterogeneous graph structure

Architecture sketch for a transformer variant that operates over multiple, related graphs simultaneously and learns the morphism between them.

Jacob Valdez
workshop submission
cite key: valdez2023multigraph

Idea

Most graph transformers operate over a single graph at a time. The generalization is unsurprising — operate over a stack of graphs that share nodes — but the interesting structure is in learning the alignment across graphs as a first-class operation, rather than treating graphs as independent inputs.

Concretely, attention is computed over a tensor indexed by (graph, node, graph, node), and a low-rank factorization keeps the cost reasonable.

Status

Draft figure done; experiments started but not finished. Filed here as a node so it doesn't fall out of the graph.

Neighborhood

Related

LLMs are the Update Rules of Intelligent Fractals: Escaping the Context Window with Iterative, Structured Local UpdatesLLMs are the Update Rul...Broadening and building beyond classical reinforcement learningBroadening and building...Research and technical writingResearch and technical ...multigraph-nnmultigraph-nnThe Node Neural Network (NNN)The Node Neural Network (NN...multi-graph-formermulti-graph-formerMulti-graph former — a transfor...