CHoE: Cross-Domain Heterogeneous Graph Prompt Learning via Structure-Conditioned Experts

Researchers introduce CHoE, a cross-domain heterogeneous graph prompt learning method that addresses the limitation of existing approaches failing when pre-training and downstream task data come from different distributions. Using structure-conditioned experts and intelligent routing mechanisms, CHoE improves performance in few-shot cross-domain applications, advancing the practical applicability of foundation models across heterogeneous graph settings.

CHoE represents a meaningful advancement in heterogeneous graph machine learning by tackling a critical gap between academic research and real-world deployment. Most existing heterogeneous graph prompt learning systems assume pre-training and application domains remain aligned, an assumption that rarely holds in production environments. This research directly addresses domain shift—a fundamental challenge in machine learning where model performance degrades when encountering data distributions different from training data.

The technical approach leverages expert networks, a proven architectural pattern from mixture-of-experts research, combined with structure-aware routing mechanisms. During pre-training, the method trains specialized experts conditioned on graph structure, then employs intelligent routing during prompt tuning to select experts that match each meta-path view's structural characteristics. The semantic fusion module integrates multi-view representations, enabling more robust predictions across diverse domains.

For the AI research community, this work strengthens foundation model transfer learning capabilities, particularly important as heterogeneous graphs appear across domains like recommendation systems, knowledge graphs, and biological networks. The few-shot learning focus addresses practical constraints where labeled data in new domains remains scarce. Industry applications benefit from models that generalize reliably across different data sources without extensive retraining.

The empirical validation showing consistent improvements over baseline approaches suggests the expert routing mechanism effectively handles structural variations between domains. Looking forward, this methodology could inspire similar approaches in other graph-based foundation models and multimodal systems where domain shift remains problematic.

• →CHoE solves domain shift problems in heterogeneous graph prompt learning through structure-conditioned expert networks and intelligent routing
• →The method achieves improved few-shot cross-domain performance compared to existing in-domain focused approaches
• →Structure-aware expert selection enables models to adapt dynamically to different graph structures across domains
• →The research bridges the gap between academic HGPL methods and real-world deployment scenarios with distribution mismatches
• →Expert-based routing with load balancing provides a scalable solution for handling multiple graph structures in foundation models