InforID: Adaptive Semantic Capacity Allocation for Parallel Generative Recommendation
A recent study published on arXiv (2608.09685) presents InforID, an innovative framework designed for adaptive semantic target construction in parallel generative recommendation. This framework tackles the challenges faced by autoregressive semantic ID recommenders, particularly the costly beam-search decoding that limits the length of item identifiers. Current approaches utilize static hyperparameters for semantic slots and codebook sizes, overlooking diverse capacity requirements. The authors demonstrate that increasing semantic slots results in minimal improvements, revealing redundancy in uniform IDs. InforID strategically distributes a fixed capacity budget among candidate slots, adjusting to different subspace requirements. This paper, which has not yet undergone peer review, is pertinent to generative recommendation systems within e-commerce and content platforms, aiming to boost both efficiency and effectiveness.
Key facts
- The paper is titled 'Adaptive Semantic Capacity Allocation for Parallel Generative Recommendation'.
- It is available on arXiv with identifier 2608.09685.
- The paper proposes InforID, a lightweight adaptive semantic target construction framework.
- InforID addresses limitations of autoregressive semantic ID recommenders, which are constrained by expensive beam-search decoding.
- Existing semantic ID methods use manually predefined and homogeneous ID structures.
- Uniformly expanding semantic slots provides limited gains, indicating redundant capacity.
- InforID allocates a fixed capacity budget across candidate semantic slots.
- The paper is a preprint and has not been peer-reviewed.
Entities
Institutions
- arXiv