The Best of the Two Worlds: Harmonizing Semantic and Hash IDs for Sequential Recommendation

Researchers propose H2Rec, a novel framework that combines Semantic IDs (SID) and Hash IDs (HID) to improve sequential recommendation systems, particularly for long-tail items. The dual-branch architecture addresses the performance trade-off between head and tail recommendations, with validation across public benchmarks and a commercial platform.

Sequential Recommender Systems face a fundamental challenge: traditional hash ID approaches excel at capturing collaborative signals for popular items but fail dramatically in long-tail scenarios where most products receive minimal interaction data. Semantic IDs offer a potential solution through multi-granular modeling and code sharing, yet they introduce a new problem—quantization mechanisms sacrifice the unique identifiers needed to distinguish head items effectively. H2Rec tackles this by architecting a dual-branch system that leverages both approaches simultaneously, maintaining the collaborative strength of hash IDs while capturing semantic richness through alternative pathways. This represents meaningful progress in recommendation science, as the long-tail problem directly impacts e-commerce and content platforms where most inventory items generate sparse engagement. The framework's dual-level alignment strategy facilitates knowledge transfer between the two representation types, creating a more robust preference model. The significance extends beyond academic interest: commercial platforms lose substantial revenue from poor long-tail recommendations, as these items typically constitute 80% of inventory but generate only 20% of transactions. Successful long-tail recommendations can unlock additional revenue streams through improved discoverability. The authors validate H2Rec across three public datasets plus live A/B testing on an unnamed large-scale platform, suggesting real-world viability. This work influences how recommendation systems balance coverage versus precision—a critical metric for platform economics. Future development likely involves integrating temporal dynamics and cross-domain knowledge transfer to further optimize the head-tail balance.

• →H2Rec combines Semantic and Hash IDs to resolve the long-standing head-tail recommendation trade-off in sequential systems.
• →Dual-branch architecture preserves unique collaborative identity for popular items while capturing multi-granular semantics for rare items.
• →Framework demonstrates consistent improvements over existing baselines in both offline and online commercial environments.
• →Long-tail recommendation quality directly impacts platform revenue since rare items constitute majority of inventory.
• →Dual-level alignment strategy enables effective knowledge transfer between heterogeneous ID representation schemes.