Enhancing Spectral Graph Recommendation via Gated Fusion and Layer Normalization

Recent advances in complementary item recommendation have highlighted the importance of modeling multi-frequency graph signals through spectral graph neural networks. While prior models like Spectral-based Complementary Graph Neural Networks (SComGNN) effectively leverage low- and mid-frequency components, their feature fusion relies on attention mechanisms, which may suffer from instability and overfitting. In this paper, we propose GSComGNN, an enhanced spectral graph recommendation model that integrates a novel GatedFusion module with Layer Normalization. The proposed GatedFusion adaptively balances the contributions from multiple graph frequency components, while LayerNorm ensures the stability of the fused representation. Moreover, we preserve the flexibility of the original model by retaining a mode selector that enables switching between attention-based and gate-based fusion. Experimental results on real-world e-commerce datasets demonstrate that our model achieves good performance in HR@K, HR@10, and NDCG metrics, validating the effectiveness and robustness of the proposed enhancements.