SGTRec: Integrating Spectral Encoding with Graph Neural Networks and Transformers for Recommendation

Recommender systems are essential for mitigating information overload in e-commerce, streaming, and social platforms. Graph neural network (GNN) recommenders exploit user–item interaction topology but can suffer from oversmoothing and local bias as depth increases. Transformer-based models capture long-range dependencies with self-attention, yet may under-utilize graph structure and become costly at scale. We propose SGTRec, a hybrid framework that combines GNN message passing with Transformer attention and injects structural context via Spectral Positional Encodings (SPE). SPEs are derived from Laplacian eigenvectors through the Graph Fourier Transform and integrate low- and high-frequency components in a unified space. To scale spectral updates across layers, we approximate spectral filtering using Chebyshev polynomials, injecting locality without repeated eigendecomposition while preserving spectral characteristics for stable learning on large interaction graphs. Experiments on multiple public benchmarks show that SGTRec consistently improves accuracy and robustness over strong baselines.