SGLP: Semantic-Guided Graph Structure Learning with LLM Structural Priors for Cross-Domain Sequential Recommendation

Abstract

Cross-Domain Sequential Recommendation (CDSR) mitigates data sparsity and cold-start issues by transferring user preferences across domains. Although Graph Neural Networks (GNNs) effectively capture collaborative signals, they struggle with the inherent structural sparsity present in individual domains. Specifically, users or items with few interactions become ''isolated nodes''. Traditional cross-domain methods fail to rescue these nodes because they rely on explicit, overlapping interaction topologies that are often missing, severely hindering message passing. In contrast, Large Language Models (LLMs) encode rich semantic knowledge yet lack the inherent ability to interpret collaborative graph structures. To address this limitation, we propose SGLP (Semantic-Guided Graph Structure Learning with LLM Structural Priors), a framework that leverages LLM knowledge as a structural prior to reconstruct sparse interaction graphs. Specifically, SGLP employs LLMs to generate Item Semantic Embeddings and User Global Semantic Profiles, establishing a robust semantic foundation. Distinct from traditional Graph Structure Learning (GSL) methods that rely solely on intrinsic node features, SGLP introduces a Semantic-Guided Graph Modeling module. This module applies differentiable structure learning to dynamically construct a task-adaptive hybrid graph, transforming unstructured semantic similarities into structured collaborative connections. To resolve distribution discrepancies between the reconstructed structural and semantic views, we enforce a Semantic-Behavioral Alignment constraint via contrastive learning. Furthermore, a Bi-Level Sequence Prediction strategy is introduced to alleviate negative transfer arising from domain shifts. Extensive experiments on three real-world Amazon datasets demonstrate that SGLP outperforms state-of-the-art baselines, validating the efficacy of integrating semantic priors into graph structure learning for cross-domain recommendation.