Semantic Structure-Aware Sequential Recommendation

Integrating item side information into ID-based sequential recommenders has proven effective in reducing popularity bias and boosting accuracy. However, among existing methods, direct fusion before sequence modeling may cause information intrusion, while indirect fusion often lacks sufficient interaction between ID and semantic embeddings or overlooks high-order semantic correlations among items. Although leveraging LLMs to generate pre-trained text embeddings has shown promise, balancing semantic and interaction information in user preference learning remains a key challenge. In this paper, we propose SSA-SR, a novel semantic structure-aware sequential recommendation model, to address the limitations of existing side information fusion methods. Firstly, we utilize text embeddings generated from a Pre-trained Language Model (PLM) to construct a KNN-connected item graph for representation learning. Upon this, we capture global semantic dependencies via spectral convolution operations and explicitly link popular and tail items, effectively avoiding information intrusion and enhancing the cold-start performance. Furthermore, we develop a dynamic attention-gated fusion module to enhance the interactions between sequential and semantic embeddings. Finally, we propose a position-wise embedding alignment strategy, and optimize the user preference representations jointly with the position-wise cross-entropy loss. Experiments on three Amazon datasets have demonstrated the superiority of SSA-SR over state-of-the-art baselines, particularly in cold-start scenarios.