Abstract

Addressing the negative transfer problem in cross-domain graph pretraining under few-shot learning scenarios, this paper proposes a multi-component pretraining framework called Graph External Attention enhanced Coordinators for Pretraining (GEA-CoPe). This framework integrates multi-head external attention with a graph coordinator. Tackling the structural and semantic discrepancies between cross-domain graphs is crucial for mitigating negative transfer; however, conventional methods often lack adaptability to complex, dynamic inter-domain variations and explicit constraints for intermediate feature distribution consistency. The proposed framework leverages an external attention-based coordinator to mediate between different graph datasets, dynamically generating cross-graph semantic alignment strategies to alleviate negative transfer induced by structural heterogeneity. It employs a dual feature normalization strategy that incorporates a cross-layer distribution alignment loss on top of intra-layer node similarity constraints, effectively suppressing feature drift. Furthermore, Kolmogorov-Arnold Networks (KAN) are introduced, whose parameter-adaptive activation functions better capture non-linear topological dependencies and enhance model interpretability. Experiments on ten real-world graph datasets demonstrate that GEA-CoPe exhibits superior cross-domain generalization capability and significantly improves performance in few-shot node classification tasks, with an average improvement of about 13.3% compared to other methods. The model can more accurately focus on critical graph structures, providing theoretical foundation and practical paradigms for deploying graph neural networks in complex scenarios.

Key words: Graph neural networks · Graph pretraining · Transfer learning · External attention