LSMRec: Leveraging Hash-Enhanced Semantic Mapping for Superior Sequential Recommendations

Traditional sequential recommendation systems often depe nd on item IDs to model user behavior sequences, which limits their ability to capture rich item semantics and adapt to new c ontexts. To address these limitations, text-based sequential rec ommendations have turned to pre-trained language models for encoding item texts. However, integrating these pre-trained se mantic representations into recommendation systems presents challenges, typically requiring extensive fine-tuning and sufferi ng from semantic misalignment. We propose the Hash-Enhanc ed Semantic Mapping Network (LSMRec), a novel framework that redefines semantic integration within recommendation sys terms. HashRec utilizes locality-sensitive hashing (LSH) to conv ert general semantic vectors from pre-trained language models into recommendation-specific representations. By mapping or iginal semantic vectors into multiple hash encoding layers and aggregating these outputs, LSMRec effectively bridges the sem antic gap, resulting in more accurate and adaptable recommen dations. Moreover, we introduce an end-to-end pre-training ta sk filter that dynamically selects and trains on multiple objecti ves, enhancing the model's ability to capture nuanced user pref erences with minimal computational overhead. Our extensive e xperiments on diverse public datasets demonstrate that HashR ec consistently outperforms state-of-the-art methods, confirmi ng its superior performance, efficiency, and robustness. This w ork advances the field of sequential recommendation and prov ides a scalable solution for integrating complex semantic infor mation into recommendation systems, setting a new standard f or future research and applications.