DyGFM: A Multi-Domain Dynamic Graph Foundation Model Using Decoupled Prompting
Researchers have proposed DyGFM, the first Dynamic Graph Foundation Model (GFM) designed to operate across multiple domains. Dynamic graphs, which model evolving relationships in systems like social networks or financial transactions, typically suffer from inconsistent semantic and temporal patterns across domains, leading to negative knowledge transfer in the standard pretrain-then-finetune paradigm. DyGFM addresses this with a dual-branch pre-training strategy that decouples transferable semantics from domain-specific dynamics, and a cross-domain divergence-conditioned prompting mechanism to mitigate negative transfer during adaptation. The work is described in a preprint on arXiv (2605.13540) and represents a step toward generalizable dynamic graph learning.
Key facts
- DyGFM is a Dynamic Graph Foundation Model for multiple domains.
- It uses a dual-branch pre-training strategy with semantic-temporal decoupling.
- A cross-domain divergence-conditioned prompting mechanism alleviates negative transfer.
- The model addresses challenges of inconsistent semantic and temporal patterns across domains.
- The work is published as a preprint on arXiv with ID 2605.13540.
- It claims to be the first multi-domain dynamic GFM.
- The pretrain-then-finetune paradigm often suffers from negative knowledge transfer.
- Dynamic graphs are ubiquitous in real-world systems.
Entities
Institutions
- arXiv