SAGERec: Semantic-Aware Global Graph-Enhanced Representation Learning for Sequential Recommendation

Sequential recommendation aims to model evolving user preferences based on historical interactions. Transformer-based architectures have achieved strong performance by focusing on user-level sequential patterns, yet global item–item relationships are often underrepresented, limiting the ability to capture broader contextual signals. In many real-world scenarios, items contain rich textual attributes such as descriptions and categories, but these semantic features are seldom exploited in existing sequential models. To address this gap, a Semantic-Aware Global Graph-Enhanced Sequential Recommendation framework (SAGERec) is developed, in which globally derived semantic structures are incorporated to enrich item representations before sequence modeling. Large language models (LLMs) are used to generate semantically grounded item embeddings, from which a global item–item graph is constructed to capture content-level relations that extend beyond behavioral co-occurrence. These semantic relations are further refined through an adaptive edge-weight learning mechanism, enabling the graph structure to align with evolving item representations during training. The adaptively enhanced item embeddings are subsequently integrated into a lightweight Transformer-based sequential encoder for next-item prediction. Extensive experiments on three benchmark datasets demonstrate that the proposed framework consistently outperforms competitive baselines, indicating that integrating LLM-derived semantics with adaptive graph refinement leads to more expressive sequential representations.