Decomposing Predictive Roles of Semantic and Collaborative Information for Sequential Recommendation

Sequential recommendation benefits from incorporating item textual semantics in addition to interaction-based collaborative signals. However, many existing methods simply concatenate or jointly encode semantic and collaborative features, without explicitly modeling how each source contributes to next-item prediction. As a result, predictive signals with fundamentally different roles are often entangled in unified representations, which limits the model's ability to exploit fine-grained predictive information. In this paper, we revisit semantics-enhanced sequential recommendation by introducing an information-theoretic analysis, which decomposes next-item predictive information into three components: shared information, semantic-unique information, and collaborative-unique information. Each component plays a distinct role as predictive evidence and contributes differently to user preference modeling. Guided by this analysis, we propose RIDRec, a novel role-aware sequential recommendation framework. RIDRec learns three role-specialized representations separately through alignment and separation constraints, each independently modeling user preferences, which preserves role distinctiveness and prevents information leakage across roles. Furthermore, RIDRec employs a role-aware fusion mechanism that adaptively combines the three role-specialized representations based on their relevance to the current user context. Extensive experiments on four public benchmarks demonstrate the effectiveness of RIDRec, showing significant improvements over various existing sequential recommendation methods. Our code is available at https://github.com/jiangnanx129/RIDRec.