Cross Regional User Preference Modeling for Agricultural E Commerce Platforms via Federated Graph Neural Networks

With the growing regionalization of agricultural e-commerce platforms, user behavior exhibits significant geographic variability. Traditional recommendation systems face challenges in generalization and privacy protection when applied to multi-platform collaborative modeling. To address issues such as data silos, cross-regional preference heterogeneity, and risks associated with centralized modeling, this paper proposes a cross-regional user preference modeling method based on Federated Graph Neural Networks (FedGNN-AE). The proposed framework preserves data locality on each platform while leveraging graph neural networks to structurally model user–item interactions. Federated learning is then employed to securely aggregate model parameters and enable global collaboration. Additionally, a semantic alignment module is introduced to perform contrastive optimization of user embeddings across platforms, improving the model's ability to adapt to regional behavioral differences. Experimental results demonstrate that FedGNN-AE outperforms existing mainstream methods in recommendation accuracy, ranking relevance, and communication efficiency, significantly enhancing recommendation performance in multi-platform scenarios. Furthermore, a simulated dataset named “AgriRec-CR” covering multiple representative regions, is constructed to serve as a reusable benchmark for future studies.