CHoE: Cross-Domain Heterogeneous Graph Prompt Learning via Structure-Conditioned Experts
The recently introduced CHoE method tackles the shortcomings of current Heterogeneous Graph Prompt Learning (HGPL) techniques, which perform well in specific domains but falter when applied to different contexts. Built on an expert network, CHoE serves as a cross-domain HGPL approach. It incorporates and trains structure-conditioned experts during the pre-training phase. For prompt tuning, it employs a mechanism for expert routing and load balancing that focuses on selecting experts that are compatible with the structure. This methodology is elaborated in a paper available on arXiv (2605.15888).
Key facts
- CHoE is a cross-domain HGPL method.
- It uses structure-conditioned experts during pre-training.
- It uses structure-aware expert routing and load balancing during prompt tuning.
- Existing HGPL methods are limited to in-domain scenarios.
- Real-world deployments often span multiple domains.
- Pre-training and downstream data may have different distributions.
- The paper is on arXiv with ID 2605.15888.
- CHoE addresses performance degradation when domains shift.
Entities
Institutions
- arXiv