Sequential recommendation aims to forecast the next item a user will interact with based on past behavior. Transformers have pushed performance forward but remain computationally expensive. Recently, State Space Models (SSMs) have shown efficiency for long-range sequences, yet most approaches assign a single static semantic role to each item, overlooking item polysemy. In practice, the same item can serve different semantic functions depending on a user's short‑term intent or long‑term preference, and these functions evolve across temporal granularities. To address this, we introduce DSRec, a dual‑interest multi‑granular SSM architecture that explicitly disentangles long‑term and short‑term item roles. The long‑term branch aggregates historical clicks into stable preference embeddings, while the short‑term branch weights local session intent by inter‑click time gaps to capture immediate needs. These embeddings are fed into distinct SSM encoders: a full‑sequence Mamba for long‑term modeling and a time‑modulated SSM that dynamically adjusts state evolution based on temporal intervals. A residual cross‑fusion mechanism aligns the two granularity streams, allowing information exchange while preserving semantic independence. Experiments on several public benchmarks compare DSRec against state‑of‑the‑art Transformers and SSM baselines, showing consistent gains in accuracy and recall.
Review