Graphlets as Structural Tokens for Knowledge Graph Foundation Models
A new framework proposes using graphlets—small connected subgraphs—as structural tokens to enable transferable representations in Knowledge Graph Foundation Models (KGFMs). Unlike language and vision, where data reduces to discrete symbols on a fixed grid, knowledge graphs have irregular, non-Euclidean topologies with varying local neighborhoods. This lack of a universal token set limits KGFMs' ability to transfer representations across unseen graphs. The paper introduces a model-agnostic approach that treats graphlets as recurring structural building blocks, closing the gap in structural vocabulary for KGFMs. The work is published on arXiv under identifier 2605.06154.
Key facts
- Knowledge Graphs lack a common fixed grid unlike language or vision.
- Graphlets are small connected graphs used as structural tokens.
- The framework is model-agnostic.
- The paper is on arXiv with ID 2605.06154.
- KGFMs rely on structural invariances for transferable representations.
- Graphlets recur in heterogeneous knowledge graphs.
- The approach addresses the lack of universal token sets for KGFMs.
- The work was announced in May 2025.
Entities
Institutions
- arXiv